Zero-Shot Classification
Transformers
Safetensors
GGUF
English
unknown
decision-model
system-one
falcondec
lightdec_v2
calibrated-decisions
multiple-choice
intent-classification
customer-support
natural-language-inference
code
guardrails
agents
selective-prediction
falconsai
model-surgeon
attested-lineage
Upload 16 files
Browse files- README.md +243 -0
- compact-int8/README.md +243 -0
- compact-int8/encoder/config.json +79 -0
- compact-int8/falcondec_config.json +54 -0
- compact-int8/falcondec_modeling.py +358 -0
- compact-int8/falcondec_report.json +1802 -0
- compact-int8/model_int8.safetensors +3 -0
- compact-int8/tokenizer/tokenizer.json +0 -0
- compact-int8/tokenizer/tokenizer_config.json +17 -0
- encoder/config.json +79 -0
- falcondec_config.json +54 -0
- falcondec_modeling.py +358 -0
- falcondec_report.json +1802 -0
- model.safetensors +3 -0
- tokenizer/tokenizer.json +0 -0
- tokenizer/tokenizer_config.json +17 -0
README.md
ADDED
|
@@ -0,0 +1,243 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
language: [en]
|
| 4 |
+
library_name: transformers
|
| 5 |
+
tags: [decision-model, system-one, calibrated-decisions, multiple-choice, zero-shot-classification, falcondec, lightdec]
|
| 6 |
+
---
|
| 7 |
+
# LightDec_V2_Long v1.0.0 — FalconDec architecture
|
| 8 |
+
|
| 9 |
+
Single-pass, typed (`choice` / `noul` / `score`), calibrated closed-set decision model. Successor to
|
| 10 |
+
[`Falconsai/proof_v2`](https://huggingface.co/Falconsai/proof_v2).
|
| 11 |
+
|
| 12 |
+
* Built with the FalconDec notebook V3.7 · backbone `jhu-clsp/ettin-encoder-150m` · mode `scratch` · preset `long` · lineage: jhu-clsp/ettin-encoder-150m
|
| 13 |
+
* Parameters: 159.7M · weights: fp16 305 MB, int8 153 MB
|
| 14 |
+
* Layout: `[CLS] question [SEP] [MASK] opt1 … [MASK] optk [SEP] state [SEP]`; set-transformer option head;
|
| 15 |
+
temperature per (question type × option-count bucket).
|
| 16 |
+
|
| 17 |
+
## Usage
|
| 18 |
+
|
| 19 |
+
```python
|
| 20 |
+
import importlib.util
|
| 21 |
+
from huggingface_hub import snapshot_download
|
| 22 |
+
path = snapshot_download("<repo>") # or a local LightDec_V2_Long-v1.0.0 directory
|
| 23 |
+
spec = importlib.util.spec_from_file_location("falcondec_modeling", f"{path}/falcondec_modeling.py")
|
| 24 |
+
fdm = importlib.util.module_from_spec(spec); spec.loader.exec_module(fdm)
|
| 25 |
+
model, tok = fdm.load_falcondec(path) # add "/compact-int8" for the int8 artefact
|
| 26 |
+
out = fdm.decide(model, tok, state={"body": "I was charged twice, refund me or I cancel."}, questions={
|
| 27 |
+
"dept": {"type": "choice", "instructions": "Which team?", "criteria": {"billing": "payments", "tech": "bugs"}},
|
| 28 |
+
"churn": {"type": "noul", "instructions": "Does the user threaten to leave?"},
|
| 29 |
+
"urgency": {"type": "score", "instructions": "How urgent?", "criteria": ["low", "medium", "high"]}})
|
| 30 |
+
```
|
| 31 |
+
|
| 32 |
+
## Test results
|
| 33 |
+
|
| 34 |
+
Overall: micro acc **0.784**, task-macro acc **0.789**, ECE **0.029**,
|
| 35 |
+
NLL 0.563, Brier 0.295, AURC 0.068.
|
| 36 |
+
|
| 37 |
+
| task | domain | heldout | n | chance | acc | proof_v2 (card) | ece |
|
| 38 |
+
|---|---|---|---|---|---|---|---|
|
| 39 |
+
| counsel/critique_quality | agentic | False | 201 | 0.333 | 0.627 | | 0.271 |
|
| 40 |
+
| agenttrek/finish_now | agentic | False | 151 | 0.500 | 0.788 | | 0.075 |
|
| 41 |
+
| counsel/step_has_error | agentic | False | 201 | 0.500 | 0.831 | | 0.148 |
|
| 42 |
+
| agenttrek/next_action_type | agentic | False | 487 | 0.191 | 0.864 | | 0.092 |
|
| 43 |
+
| hotpotqa/retrieve | agentic | False | 497 | 0.167 | 0.893 | | 0.043 |
|
| 44 |
+
| hotpotqa/comparison_yes_no | agentic | False | 26 | 0.500 | 0.923 | | 0.082 |
|
| 45 |
+
| sst5/score | classification | True | 500 | 0.200 | 0.406 | | 0.069 |
|
| 46 |
+
| emotion/6way | classification | True | 500 | 0.167 | 0.472 | | 0.211 |
|
| 47 |
+
| yelp/score | classification | False | 500 | 0.200 | 0.668 | | 0.091 |
|
| 48 |
+
| ag_news/topic | classification | False | 500 | 0.250 | 0.900 | | 0.052 |
|
| 49 |
+
| devign/vulnerability | code | False | 500 | 0.500 | 0.632 | 0.542 | 0.070 |
|
| 50 |
+
| humaneval/completion | code | True | 119 | 0.394 | 0.849 | 0.575 | 0.140 |
|
| 51 |
+
| mbpp/bugspot | code | False | 256 | 0.406 | 0.855 | 0.475 | 0.037 |
|
| 52 |
+
| codexglue/func_name | code | False | 467 | 0.255 | 0.955 | 0.901 | 0.039 |
|
| 53 |
+
| bigclonebench/clone | code | False | 500 | 0.500 | 0.962 | 0.383 | 0.031 |
|
| 54 |
+
| codexglue/doc_to_code | code | False | 504 | 0.279 | 0.976 | 0.961 | 0.013 |
|
| 55 |
+
| mbpp/solution | code | False | 500 | 0.250 | 0.978 | 0.992 | 0.020 |
|
| 56 |
+
| codexglue/code_to_doc | code | False | 504 | 0.262 | 0.990 | 0.969 | 0.010 |
|
| 57 |
+
| codexglue/lang_id | code | False | 504 | 0.235 | 1.000 | 0.997 | 0.001 |
|
| 58 |
+
| prompt_injections/detect | guardrails | True | 116 | 0.500 | 0.595 | | 0.333 |
|
| 59 |
+
| agentharm/refuse | guardrails | True | 416 | 0.500 | 0.700 | | 0.089 |
|
| 60 |
+
| civil_comments/toxic | guardrails | False | 500 | 0.500 | 0.926 | | 0.051 |
|
| 61 |
+
| jailbreak/detect | guardrails | False | 262 | 0.500 | 0.973 | | 0.031 |
|
| 62 |
+
| banking77/intent_77 | intents | True | 300 | 0.013 | 0.580 | | 0.241 |
|
| 63 |
+
| banking77/intent | intents | True | 500 | 0.321 | 0.922 | 0.883 | 0.037 |
|
| 64 |
+
| massive_en/intent | intents | False | 500 | 0.130 | 0.930 | | 0.028 |
|
| 65 |
+
| clinc150/intent | intents | False | 500 | 0.124 | 0.970 | 0.850 | 0.015 |
|
| 66 |
+
| long/contract_clause | long_context | False | 150 | 0.200 | 1.000 | | 0.000 |
|
| 67 |
+
| long/email_thread | long_context | False | 150 | 0.250 | 1.000 | | 0.000 |
|
| 68 |
+
| long/service_log | long_context | False | 150 | 0.200 | 1.000 | | 0.001 |
|
| 69 |
+
| policy/table_count_transfer | policy | False | 500 | 0.121 | 0.324 | | 0.541 |
|
| 70 |
+
| policy/invoice_total_transfer | policy | False | 500 | 0.500 | 0.474 | | 0.082 |
|
| 71 |
+
| policy/sla_urgency_transfer | policy | False | 500 | 0.250 | 0.606 | | 0.131 |
|
| 72 |
+
| policy/invoice_overdue_transfer | policy | False | 500 | 0.500 | 0.824 | | 0.020 |
|
| 73 |
+
| policy/count_threshold_transfer | policy | False | 500 | 0.105 | 0.880 | | 0.048 |
|
| 74 |
+
| policy/table_compare_transfer | policy | False | 500 | 0.500 | 0.928 | | 0.049 |
|
| 75 |
+
| policy/table_extreme_transfer | policy | False | 500 | 0.122 | 0.932 | | 0.053 |
|
| 76 |
+
| policy/free_shipping_transfer | policy | False | 500 | 0.500 | 0.938 | | 0.041 |
|
| 77 |
+
| policy/refund_approval_transfer | policy | False | 500 | 0.333 | 0.962 | | 0.019 |
|
| 78 |
+
| policy/access_control_transfer | policy | False | 500 | 0.333 | 1.000 | | 0.002 |
|
| 79 |
+
| policy/return_window_transfer | policy | False | 500 | 0.333 | 1.000 | | 0.026 |
|
| 80 |
+
| aqua_rat/math | reasoning | False | 247 | 0.200 | 0.340 | | 0.080 |
|
| 81 |
+
| mmlu/mcq | reasoning | True | 500 | 0.250 | 0.390 | | 0.155 |
|
| 82 |
+
| anli/nli | reasoning | False | 498 | 0.333 | 0.486 | | 0.146 |
|
| 83 |
+
| arc_challenge/mcq | reasoning | True | 500 | 0.250 | 0.490 | 0.308 | 0.168 |
|
| 84 |
+
| openbookqa/mcq | reasoning | False | 500 | 0.250 | 0.572 | 0.292 | 0.218 |
|
| 85 |
+
| hellaswag/continuation | reasoning | False | 500 | 0.250 | 0.576 | | 0.075 |
|
| 86 |
+
| arc_easy/mcq | reasoning | True | 500 | 0.250 | 0.612 | 0.425 | 0.122 |
|
| 87 |
+
| commonsense_qa/mcq | reasoning | False | 493 | 0.200 | 0.643 | 0.442 | 0.143 |
|
| 88 |
+
| winogrande/blank | reasoning | False | 500 | 0.500 | 0.662 | | 0.116 |
|
| 89 |
+
| gsm8k/math | reasoning | False | 500 | 0.250 | 0.700 | 0.275 | 0.085 |
|
| 90 |
+
| boolq/yes_no | reasoning | False | 500 | 0.500 | 0.822 | 0.717 | 0.080 |
|
| 91 |
+
| mnli/claim | reasoning | False | 500 | 0.333 | 0.860 | 0.492 | 0.081 |
|
| 92 |
+
| snli/nli | reasoning | False | 988 | 0.333 | 0.895 | | 0.108 |
|
| 93 |
+
| sciq/mcq | reasoning | False | 498 | 0.250 | 0.954 | 0.692 | 0.021 |
|
| 94 |
+
| scitail/support | reasoning | False | 500 | 0.500 | 0.962 | | 0.030 |
|
| 95 |
+
| qasc/mcq | reasoning | False | 500 | 0.125 | 0.986 | | 0.008 |
|
| 96 |
+
| snli/must_be_true | reasoning | False | 500 | 0.333 | 0.986 | 0.908 | 0.043 |
|
| 97 |
+
| snli/contradicts | reasoning | False | 500 | 0.333 | 0.990 | 0.892 | 0.019 |
|
| 98 |
+
| triage/support_email | support | False | 505 | 0.323 | 0.945 | | 0.046 |
|
| 99 |
+
| bitext/category | support | False | 500 | 0.161 | 1.000 | | 0.001 |
|
| 100 |
+
| bitext/route | support | False | 500 | 0.190 | 1.000 | 0.958 | 0.000 |
|
| 101 |
+
| tev1_test/sst5 | tev1_benchmark | True | 150 | 0.200 | 0.393 | | 0.108 |
|
| 102 |
+
| tev1_test/routing | tev1_benchmark | True | 600 | 0.200 | 0.458 | | 0.038 |
|
| 103 |
+
| tev1_test/policy | tev1_benchmark | True | 1200 | 0.333 | 0.529 | | 0.059 |
|
| 104 |
+
| tev1_test/banking77 | tev1_benchmark | True | 200 | 0.178 | 0.720 | | 0.137 |
|
| 105 |
+
| tev1_test/mnli | tev1_benchmark | True | 300 | 0.333 | 0.723 | | 0.075 |
|
| 106 |
+
| tev1_test/boolq | tev1_benchmark | True | 200 | 0.500 | 0.835 | | 0.050 |
|
| 107 |
+
| tev1_test/ag_news | tev1_benchmark | True | 150 | 0.250 | 0.920 | | 0.098 |
|
| 108 |
+
| typed_decisions/agent_trace_observability | workflows | False | 500 | 0.300 | 0.734 | | 0.222 |
|
| 109 |
+
| typed_decisions/customer_service | workflows | False | 500 | 0.280 | 0.764 | | 0.210 |
|
| 110 |
+
| typed_decisions/security_incidents | workflows | False | 500 | 0.340 | 0.768 | | 0.235 |
|
| 111 |
+
| typed_decisions/invoice_processing | workflows | False | 500 | 0.350 | 0.826 | | 0.195 |
|
| 112 |
+
|
| 113 |
+
|
| 114 |
+
### Head-to-head with proof_v2 on identical test decisions
|
| 115 |
+
|
| 116 |
+
| task | n | LightDec | proof_v2 | Δ |
|
| 117 |
+
|---|---|---|---|---|
|
| 118 |
+
| ag_news/topic | 120.000 | 0.850 | 0.450 | 0.400 |
|
| 119 |
+
| agentharm/refuse | 120.000 | 0.667 | 0.450 | 0.217 |
|
| 120 |
+
| agenttrek/finish_now | 120.000 | 0.783 | 0.258 | 0.525 |
|
| 121 |
+
| agenttrek/next_action_type | 120.000 | 0.875 | 0.075 | 0.800 |
|
| 122 |
+
| anli/nli | 120.000 | 0.592 | 0.325 | 0.267 |
|
| 123 |
+
| aqua_rat/math | 120.000 | 0.350 | 0.242 | 0.108 |
|
| 124 |
+
| arc_challenge/mcq | 120.000 | 0.533 | 0.308 | 0.225 |
|
| 125 |
+
| arc_easy/mcq | 120.000 | 0.650 | 0.458 | 0.192 |
|
| 126 |
+
| banking77/intent | 120.000 | 0.925 | 0.867 | 0.058 |
|
| 127 |
+
| bigclonebench/clone | 120.000 | 0.967 | 0.242 | 0.725 |
|
| 128 |
+
| bitext/category | 120.000 | 1.000 | 0.858 | 0.142 |
|
| 129 |
+
| bitext/route | 120.000 | 1.000 | 0.925 | 0.075 |
|
| 130 |
+
| boolq/yes_no | 120.000 | 0.825 | 0.667 | 0.158 |
|
| 131 |
+
| civil_comments/toxic | 120.000 | 0.933 | 0.358 | 0.575 |
|
| 132 |
+
| clinc150/intent | 120.000 | 0.975 | 0.642 | 0.333 |
|
| 133 |
+
| codexglue/code_to_doc | 120.000 | 1.000 | 0.933 | 0.067 |
|
| 134 |
+
| codexglue/doc_to_code | 120.000 | 0.983 | 0.975 | 0.008 |
|
| 135 |
+
| codexglue/func_name | 120.000 | 0.992 | 0.917 | 0.075 |
|
| 136 |
+
| codexglue/lang_id | 120.000 | 1.000 | 1.000 | 0.000 |
|
| 137 |
+
| commonsense_qa/mcq | 120.000 | 0.683 | 0.408 | 0.275 |
|
| 138 |
+
| counsel/critique_quality | 120.000 | 0.608 | 0.258 | 0.350 |
|
| 139 |
+
| counsel/step_has_error | 120.000 | 0.825 | 0.750 | 0.075 |
|
| 140 |
+
| devign/vulnerability | 120.000 | 0.667 | 0.567 | 0.100 |
|
| 141 |
+
| emotion/6way | 120.000 | 0.492 | 0.367 | 0.125 |
|
| 142 |
+
| gsm8k/math | 120.000 | 0.733 | 0.125 | 0.608 |
|
| 143 |
+
| hellaswag/continuation | 120.000 | 0.608 | 0.292 | 0.317 |
|
| 144 |
+
| hotpotqa/comparison_yes_no | 26.000 | 0.923 | 0.462 | 0.462 |
|
| 145 |
+
| hotpotqa/retrieve | 120.000 | 0.875 | 0.292 | 0.583 |
|
| 146 |
+
| humaneval/completion | 119.000 | 0.849 | 0.513 | 0.336 |
|
| 147 |
+
| jailbreak/detect | 120.000 | 0.983 | 0.567 | 0.417 |
|
| 148 |
+
| long/contract_clause | 120.000 | 1.000 | 0.233 | 0.767 |
|
| 149 |
+
| long/email_thread | 120.000 | 1.000 | 0.050 | 0.950 |
|
| 150 |
+
| long/service_log | 120.000 | 1.000 | 0.242 | 0.758 |
|
| 151 |
+
| massive_en/intent | 120.000 | 0.950 | 0.742 | 0.208 |
|
| 152 |
+
| mbpp/bugspot | 120.000 | 0.858 | 0.575 | 0.283 |
|
| 153 |
+
| mbpp/solution | 120.000 | 1.000 | 0.850 | 0.150 |
|
| 154 |
+
| mmlu/mcq | 120.000 | 0.408 | 0.308 | 0.100 |
|
| 155 |
+
| mnli/claim | 120.000 | 0.875 | 0.317 | 0.558 |
|
| 156 |
+
| openbookqa/mcq | 120.000 | 0.650 | 0.283 | 0.367 |
|
| 157 |
+
| policy/access_control_transfer | 120.000 | 1.000 | 0.233 | 0.767 |
|
| 158 |
+
| policy/count_threshold_transfer | 120.000 | 0.867 | 0.108 | 0.758 |
|
| 159 |
+
| policy/free_shipping_transfer | 120.000 | 0.925 | 0.708 | 0.217 |
|
| 160 |
+
| policy/invoice_overdue_transfer | 120.000 | 0.808 | 0.658 | 0.150 |
|
| 161 |
+
| policy/invoice_total_transfer | 120.000 | 0.450 | 0.525 | -0.075 |
|
| 162 |
+
| policy/refund_approval_transfer | 120.000 | 0.958 | 0.325 | 0.633 |
|
| 163 |
+
| policy/return_window_transfer | 120.000 | 1.000 | 0.342 | 0.658 |
|
| 164 |
+
| policy/sla_urgency_transfer | 120.000 | 0.517 | 0.442 | 0.075 |
|
| 165 |
+
| policy/table_compare_transfer | 120.000 | 0.917 | 0.492 | 0.425 |
|
| 166 |
+
| policy/table_count_transfer | 120.000 | 0.317 | 0.158 | 0.158 |
|
| 167 |
+
| policy/table_extreme_transfer | 120.000 | 0.925 | 0.142 | 0.783 |
|
| 168 |
+
| prompt_injections/detect | 116.000 | 0.595 | 0.543 | 0.052 |
|
| 169 |
+
| qasc/mcq | 120.000 | 0.992 | 0.858 | 0.133 |
|
| 170 |
+
| sciq/mcq | 120.000 | 0.967 | 0.867 | 0.100 |
|
| 171 |
+
| scitail/support | 120.000 | 0.958 | 0.425 | 0.533 |
|
| 172 |
+
| snli/contradicts | 120.000 | 1.000 | 0.358 | 0.642 |
|
| 173 |
+
| snli/must_be_true | 120.000 | 0.992 | 0.925 | 0.067 |
|
| 174 |
+
| snli/nli | 120.000 | 0.908 | 0.342 | 0.567 |
|
| 175 |
+
| sst5/score | 120.000 | 0.383 | 0.283 | 0.100 |
|
| 176 |
+
| tev1_test/ag_news | 120.000 | 0.925 | 0.425 | 0.500 |
|
| 177 |
+
| tev1_test/banking77 | 120.000 | 0.717 | 0.550 | 0.167 |
|
| 178 |
+
| tev1_test/boolq | 120.000 | 0.842 | 0.558 | 0.283 |
|
| 179 |
+
| tev1_test/mnli | 120.000 | 0.750 | 0.392 | 0.358 |
|
| 180 |
+
| tev1_test/policy | 120.000 | 0.500 | 0.408 | 0.092 |
|
| 181 |
+
| tev1_test/routing | 120.000 | 0.442 | 0.292 | 0.150 |
|
| 182 |
+
| tev1_test/sst5 | 120.000 | 0.392 | 0.275 | 0.117 |
|
| 183 |
+
| triage/support_email | 120.000 | 0.958 | 0.392 | 0.567 |
|
| 184 |
+
| typed_decisions/agent_trace_observability | 120.000 | 0.808 | 0.333 | 0.475 |
|
| 185 |
+
| typed_decisions/customer_service | 120.000 | 0.742 | 0.283 | 0.458 |
|
| 186 |
+
| typed_decisions/invoice_processing | 120.000 | 0.817 | 0.475 | 0.342 |
|
| 187 |
+
| typed_decisions/security_incidents | 120.000 | 0.758 | 0.500 | 0.258 |
|
| 188 |
+
| winogrande/blank | 120.000 | 0.692 | 0.558 | 0.133 |
|
| 189 |
+
| yelp/score | 120.000 | 0.625 | 0.325 | 0.300 |
|
| 190 |
+
|
| 191 |
+
### Tev1 benchmark (tev1 v2.1 test records, held out)
|
| 192 |
+
|
| 193 |
+
| task | n | accuracy | source also in LightDec training |
|
| 194 |
+
|---|---|---|---|
|
| 195 |
+
| tev1_test/ag_news | 150 | 0.920 | yes |
|
| 196 |
+
| tev1_test/banking77 | 200 | 0.720 | no |
|
| 197 |
+
| tev1_test/boolq | 200 | 0.835 | yes |
|
| 198 |
+
| tev1_test/mnli | 300 | 0.723 | yes |
|
| 199 |
+
| tev1_test/policy | 1200 | 0.529 | no |
|
| 200 |
+
| tev1_test/routing | 600 | 0.458 | no |
|
| 201 |
+
| tev1_test/sst5 | 150 | 0.393 | no |
|
| 202 |
+
|
| 203 |
+
All categories 0.584 on 2800 decisions; excluding sources LightDec also trains on 0.518. Tev1-4B's published results (880/1,000 main decisions,
|
| 204 |
+
300/300 policy transfer) are reused development benchmarks and may not be the same records.
|
| 205 |
+
|
| 206 |
+
### Support-email triage (`triage/support_email`, 101 held-out emails)
|
| 207 |
+
|
| 208 |
+
Per question: topic 0.990 · refund 1.000 · breakage 0.990 · anger 0.921 · judgment 0.822. Routing-pile accuracy (person / engineering / feature log / reply): 0.941.
|
| 209 |
+
|
| 210 |
+
### Long-state tasks at max_len 2048
|
| 211 |
+
|
| 212 |
+
long/contract_clause 1.000 (chance 0.20) · long/email_thread 1.000 (chance 0.25) · long/service_log 1.000 (chance 0.20)
|
| 213 |
+
|
| 214 |
+
### Latency (measured on NVIDIA RTX PRO 6000 Blackwell Server Edition)
|
| 215 |
+
|
| 216 |
+
| device | weights | questions per call | p50 ms | p95 ms |
|
| 217 |
+
|---|---|---|---|---|
|
| 218 |
+
| cuda | fp16 | 1 | 10.1 | 10.4 |
|
| 219 |
+
| cuda | fp16 | 5 | 11.3 | 11.4 |
|
| 220 |
+
| cuda | fp16 | 10 | 12.6 | 12.7 |
|
| 221 |
+
| cpu fp32 | int8 file | 1 | 48.4 | 48.8 |
|
| 222 |
+
|
| 223 |
+
Batched throughput: 2589 decisions/s.
|
| 224 |
+
|
| 225 |
+
## Training
|
| 226 |
+
|
| 227 |
+
Strictly proper objective (log + 0.5·spherical + 1.0·RPS for ordinal), soft teacher targets where available,
|
| 228 |
+
task-balanced sampling (α = 0.5), NOTA augmentation (8%), per-epoch option reshuffling,
|
| 229 |
+
AdamW + LLRD 0.9, EMA 0.999, 10 epoch(s), RLCD stage off/reverted.
|
| 230 |
+
Training time 291.7 min on NVIDIA RTX PRO 6000 Blackwell Server Edition.
|
| 231 |
+
|
| 232 |
+
## Limitations
|
| 233 |
+
|
| 234 |
+
* Closed world: the model ranks the options you give it. Add "None of the above" when that is a valid answer (it was trained with it).
|
| 235 |
+
* Multi-step arithmetic (GSM8K / AQuA) is a System-2 task; route low-confidence answers to an LLM (`defer` flag, threshold 0.7).
|
| 236 |
+
* English only; code judgement (vulnerabilities, clones) is still weak; check calibration on your own traffic.
|
| 237 |
+
|
| 238 |
+
## References
|
| 239 |
+
|
| 240 |
+
Warner et al. 2024 (ModernBERT) · Weller et al. 2025 (Ettin) · Gneiting & Raftery 2007 (strictly proper scoring rules) ·
|
| 241 |
+
Guo et al. 2017 (temperature scaling) · Geifman & El-Yaniv 2017 (selective classification) · Shao et al. 2024 (GRPO) ·
|
| 242 |
+
Williams 1992 (REINFORCE) · Zaheer et al. 2017 / Lee et al. 2019 (Deep Sets / Set Transformer) · Hinton et al. 2015 (distillation) ·
|
| 243 |
+
Laya (convaiinnovations) · TypeSafe Jev · Together Tev1.
|
compact-int8/README.md
ADDED
|
@@ -0,0 +1,243 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
language: [en]
|
| 4 |
+
library_name: transformers
|
| 5 |
+
tags: [decision-model, system-one, calibrated-decisions, multiple-choice, zero-shot-classification, falcondec, lightdec]
|
| 6 |
+
---
|
| 7 |
+
# LightDec_V2_Long v1.0.0 — FalconDec architecture
|
| 8 |
+
|
| 9 |
+
Single-pass, typed (`choice` / `noul` / `score`), calibrated closed-set decision model. Successor to
|
| 10 |
+
[`Falconsai/proof_v2`](https://huggingface.co/Falconsai/proof_v2).
|
| 11 |
+
|
| 12 |
+
* Built with the FalconDec notebook V3.7 · backbone `jhu-clsp/ettin-encoder-150m` · mode `scratch` · preset `long` · lineage: jhu-clsp/ettin-encoder-150m
|
| 13 |
+
* Parameters: 159.7M · weights: fp16 305 MB, int8 153 MB
|
| 14 |
+
* Layout: `[CLS] question [SEP] [MASK] opt1 … [MASK] optk [SEP] state [SEP]`; set-transformer option head;
|
| 15 |
+
temperature per (question type × option-count bucket).
|
| 16 |
+
|
| 17 |
+
## Usage
|
| 18 |
+
|
| 19 |
+
```python
|
| 20 |
+
import importlib.util
|
| 21 |
+
from huggingface_hub import snapshot_download
|
| 22 |
+
path = snapshot_download("<repo>") # or a local LightDec_V2_Long-v1.0.0 directory
|
| 23 |
+
spec = importlib.util.spec_from_file_location("falcondec_modeling", f"{path}/falcondec_modeling.py")
|
| 24 |
+
fdm = importlib.util.module_from_spec(spec); spec.loader.exec_module(fdm)
|
| 25 |
+
model, tok = fdm.load_falcondec(path) # add "/compact-int8" for the int8 artefact
|
| 26 |
+
out = fdm.decide(model, tok, state={"body": "I was charged twice, refund me or I cancel."}, questions={
|
| 27 |
+
"dept": {"type": "choice", "instructions": "Which team?", "criteria": {"billing": "payments", "tech": "bugs"}},
|
| 28 |
+
"churn": {"type": "noul", "instructions": "Does the user threaten to leave?"},
|
| 29 |
+
"urgency": {"type": "score", "instructions": "How urgent?", "criteria": ["low", "medium", "high"]}})
|
| 30 |
+
```
|
| 31 |
+
|
| 32 |
+
## Test results
|
| 33 |
+
|
| 34 |
+
Overall: micro acc **0.784**, task-macro acc **0.789**, ECE **0.029**,
|
| 35 |
+
NLL 0.563, Brier 0.295, AURC 0.068.
|
| 36 |
+
|
| 37 |
+
| task | domain | heldout | n | chance | acc | proof_v2 (card) | ece |
|
| 38 |
+
|---|---|---|---|---|---|---|---|
|
| 39 |
+
| counsel/critique_quality | agentic | False | 201 | 0.333 | 0.627 | | 0.271 |
|
| 40 |
+
| agenttrek/finish_now | agentic | False | 151 | 0.500 | 0.788 | | 0.075 |
|
| 41 |
+
| counsel/step_has_error | agentic | False | 201 | 0.500 | 0.831 | | 0.148 |
|
| 42 |
+
| agenttrek/next_action_type | agentic | False | 487 | 0.191 | 0.864 | | 0.092 |
|
| 43 |
+
| hotpotqa/retrieve | agentic | False | 497 | 0.167 | 0.893 | | 0.043 |
|
| 44 |
+
| hotpotqa/comparison_yes_no | agentic | False | 26 | 0.500 | 0.923 | | 0.082 |
|
| 45 |
+
| sst5/score | classification | True | 500 | 0.200 | 0.406 | | 0.069 |
|
| 46 |
+
| emotion/6way | classification | True | 500 | 0.167 | 0.472 | | 0.211 |
|
| 47 |
+
| yelp/score | classification | False | 500 | 0.200 | 0.668 | | 0.091 |
|
| 48 |
+
| ag_news/topic | classification | False | 500 | 0.250 | 0.900 | | 0.052 |
|
| 49 |
+
| devign/vulnerability | code | False | 500 | 0.500 | 0.632 | 0.542 | 0.070 |
|
| 50 |
+
| humaneval/completion | code | True | 119 | 0.394 | 0.849 | 0.575 | 0.140 |
|
| 51 |
+
| mbpp/bugspot | code | False | 256 | 0.406 | 0.855 | 0.475 | 0.037 |
|
| 52 |
+
| codexglue/func_name | code | False | 467 | 0.255 | 0.955 | 0.901 | 0.039 |
|
| 53 |
+
| bigclonebench/clone | code | False | 500 | 0.500 | 0.962 | 0.383 | 0.031 |
|
| 54 |
+
| codexglue/doc_to_code | code | False | 504 | 0.279 | 0.976 | 0.961 | 0.013 |
|
| 55 |
+
| mbpp/solution | code | False | 500 | 0.250 | 0.978 | 0.992 | 0.020 |
|
| 56 |
+
| codexglue/code_to_doc | code | False | 504 | 0.262 | 0.990 | 0.969 | 0.010 |
|
| 57 |
+
| codexglue/lang_id | code | False | 504 | 0.235 | 1.000 | 0.997 | 0.001 |
|
| 58 |
+
| prompt_injections/detect | guardrails | True | 116 | 0.500 | 0.595 | | 0.333 |
|
| 59 |
+
| agentharm/refuse | guardrails | True | 416 | 0.500 | 0.700 | | 0.089 |
|
| 60 |
+
| civil_comments/toxic | guardrails | False | 500 | 0.500 | 0.926 | | 0.051 |
|
| 61 |
+
| jailbreak/detect | guardrails | False | 262 | 0.500 | 0.973 | | 0.031 |
|
| 62 |
+
| banking77/intent_77 | intents | True | 300 | 0.013 | 0.580 | | 0.241 |
|
| 63 |
+
| banking77/intent | intents | True | 500 | 0.321 | 0.922 | 0.883 | 0.037 |
|
| 64 |
+
| massive_en/intent | intents | False | 500 | 0.130 | 0.930 | | 0.028 |
|
| 65 |
+
| clinc150/intent | intents | False | 500 | 0.124 | 0.970 | 0.850 | 0.015 |
|
| 66 |
+
| long/contract_clause | long_context | False | 150 | 0.200 | 1.000 | | 0.000 |
|
| 67 |
+
| long/email_thread | long_context | False | 150 | 0.250 | 1.000 | | 0.000 |
|
| 68 |
+
| long/service_log | long_context | False | 150 | 0.200 | 1.000 | | 0.001 |
|
| 69 |
+
| policy/table_count_transfer | policy | False | 500 | 0.121 | 0.324 | | 0.541 |
|
| 70 |
+
| policy/invoice_total_transfer | policy | False | 500 | 0.500 | 0.474 | | 0.082 |
|
| 71 |
+
| policy/sla_urgency_transfer | policy | False | 500 | 0.250 | 0.606 | | 0.131 |
|
| 72 |
+
| policy/invoice_overdue_transfer | policy | False | 500 | 0.500 | 0.824 | | 0.020 |
|
| 73 |
+
| policy/count_threshold_transfer | policy | False | 500 | 0.105 | 0.880 | | 0.048 |
|
| 74 |
+
| policy/table_compare_transfer | policy | False | 500 | 0.500 | 0.928 | | 0.049 |
|
| 75 |
+
| policy/table_extreme_transfer | policy | False | 500 | 0.122 | 0.932 | | 0.053 |
|
| 76 |
+
| policy/free_shipping_transfer | policy | False | 500 | 0.500 | 0.938 | | 0.041 |
|
| 77 |
+
| policy/refund_approval_transfer | policy | False | 500 | 0.333 | 0.962 | | 0.019 |
|
| 78 |
+
| policy/access_control_transfer | policy | False | 500 | 0.333 | 1.000 | | 0.002 |
|
| 79 |
+
| policy/return_window_transfer | policy | False | 500 | 0.333 | 1.000 | | 0.026 |
|
| 80 |
+
| aqua_rat/math | reasoning | False | 247 | 0.200 | 0.340 | | 0.080 |
|
| 81 |
+
| mmlu/mcq | reasoning | True | 500 | 0.250 | 0.390 | | 0.155 |
|
| 82 |
+
| anli/nli | reasoning | False | 498 | 0.333 | 0.486 | | 0.146 |
|
| 83 |
+
| arc_challenge/mcq | reasoning | True | 500 | 0.250 | 0.490 | 0.308 | 0.168 |
|
| 84 |
+
| openbookqa/mcq | reasoning | False | 500 | 0.250 | 0.572 | 0.292 | 0.218 |
|
| 85 |
+
| hellaswag/continuation | reasoning | False | 500 | 0.250 | 0.576 | | 0.075 |
|
| 86 |
+
| arc_easy/mcq | reasoning | True | 500 | 0.250 | 0.612 | 0.425 | 0.122 |
|
| 87 |
+
| commonsense_qa/mcq | reasoning | False | 493 | 0.200 | 0.643 | 0.442 | 0.143 |
|
| 88 |
+
| winogrande/blank | reasoning | False | 500 | 0.500 | 0.662 | | 0.116 |
|
| 89 |
+
| gsm8k/math | reasoning | False | 500 | 0.250 | 0.700 | 0.275 | 0.085 |
|
| 90 |
+
| boolq/yes_no | reasoning | False | 500 | 0.500 | 0.822 | 0.717 | 0.080 |
|
| 91 |
+
| mnli/claim | reasoning | False | 500 | 0.333 | 0.860 | 0.492 | 0.081 |
|
| 92 |
+
| snli/nli | reasoning | False | 988 | 0.333 | 0.895 | | 0.108 |
|
| 93 |
+
| sciq/mcq | reasoning | False | 498 | 0.250 | 0.954 | 0.692 | 0.021 |
|
| 94 |
+
| scitail/support | reasoning | False | 500 | 0.500 | 0.962 | | 0.030 |
|
| 95 |
+
| qasc/mcq | reasoning | False | 500 | 0.125 | 0.986 | | 0.008 |
|
| 96 |
+
| snli/must_be_true | reasoning | False | 500 | 0.333 | 0.986 | 0.908 | 0.043 |
|
| 97 |
+
| snli/contradicts | reasoning | False | 500 | 0.333 | 0.990 | 0.892 | 0.019 |
|
| 98 |
+
| triage/support_email | support | False | 505 | 0.323 | 0.945 | | 0.046 |
|
| 99 |
+
| bitext/category | support | False | 500 | 0.161 | 1.000 | | 0.001 |
|
| 100 |
+
| bitext/route | support | False | 500 | 0.190 | 1.000 | 0.958 | 0.000 |
|
| 101 |
+
| tev1_test/sst5 | tev1_benchmark | True | 150 | 0.200 | 0.393 | | 0.108 |
|
| 102 |
+
| tev1_test/routing | tev1_benchmark | True | 600 | 0.200 | 0.458 | | 0.038 |
|
| 103 |
+
| tev1_test/policy | tev1_benchmark | True | 1200 | 0.333 | 0.529 | | 0.059 |
|
| 104 |
+
| tev1_test/banking77 | tev1_benchmark | True | 200 | 0.178 | 0.720 | | 0.137 |
|
| 105 |
+
| tev1_test/mnli | tev1_benchmark | True | 300 | 0.333 | 0.723 | | 0.075 |
|
| 106 |
+
| tev1_test/boolq | tev1_benchmark | True | 200 | 0.500 | 0.835 | | 0.050 |
|
| 107 |
+
| tev1_test/ag_news | tev1_benchmark | True | 150 | 0.250 | 0.920 | | 0.098 |
|
| 108 |
+
| typed_decisions/agent_trace_observability | workflows | False | 500 | 0.300 | 0.734 | | 0.222 |
|
| 109 |
+
| typed_decisions/customer_service | workflows | False | 500 | 0.280 | 0.764 | | 0.210 |
|
| 110 |
+
| typed_decisions/security_incidents | workflows | False | 500 | 0.340 | 0.768 | | 0.235 |
|
| 111 |
+
| typed_decisions/invoice_processing | workflows | False | 500 | 0.350 | 0.826 | | 0.195 |
|
| 112 |
+
|
| 113 |
+
|
| 114 |
+
### Head-to-head with proof_v2 on identical test decisions
|
| 115 |
+
|
| 116 |
+
| task | n | LightDec | proof_v2 | Δ |
|
| 117 |
+
|---|---|---|---|---|
|
| 118 |
+
| ag_news/topic | 120.000 | 0.850 | 0.450 | 0.400 |
|
| 119 |
+
| agentharm/refuse | 120.000 | 0.667 | 0.450 | 0.217 |
|
| 120 |
+
| agenttrek/finish_now | 120.000 | 0.783 | 0.258 | 0.525 |
|
| 121 |
+
| agenttrek/next_action_type | 120.000 | 0.875 | 0.075 | 0.800 |
|
| 122 |
+
| anli/nli | 120.000 | 0.592 | 0.325 | 0.267 |
|
| 123 |
+
| aqua_rat/math | 120.000 | 0.350 | 0.242 | 0.108 |
|
| 124 |
+
| arc_challenge/mcq | 120.000 | 0.533 | 0.308 | 0.225 |
|
| 125 |
+
| arc_easy/mcq | 120.000 | 0.650 | 0.458 | 0.192 |
|
| 126 |
+
| banking77/intent | 120.000 | 0.925 | 0.867 | 0.058 |
|
| 127 |
+
| bigclonebench/clone | 120.000 | 0.967 | 0.242 | 0.725 |
|
| 128 |
+
| bitext/category | 120.000 | 1.000 | 0.858 | 0.142 |
|
| 129 |
+
| bitext/route | 120.000 | 1.000 | 0.925 | 0.075 |
|
| 130 |
+
| boolq/yes_no | 120.000 | 0.825 | 0.667 | 0.158 |
|
| 131 |
+
| civil_comments/toxic | 120.000 | 0.933 | 0.358 | 0.575 |
|
| 132 |
+
| clinc150/intent | 120.000 | 0.975 | 0.642 | 0.333 |
|
| 133 |
+
| codexglue/code_to_doc | 120.000 | 1.000 | 0.933 | 0.067 |
|
| 134 |
+
| codexglue/doc_to_code | 120.000 | 0.983 | 0.975 | 0.008 |
|
| 135 |
+
| codexglue/func_name | 120.000 | 0.992 | 0.917 | 0.075 |
|
| 136 |
+
| codexglue/lang_id | 120.000 | 1.000 | 1.000 | 0.000 |
|
| 137 |
+
| commonsense_qa/mcq | 120.000 | 0.683 | 0.408 | 0.275 |
|
| 138 |
+
| counsel/critique_quality | 120.000 | 0.608 | 0.258 | 0.350 |
|
| 139 |
+
| counsel/step_has_error | 120.000 | 0.825 | 0.750 | 0.075 |
|
| 140 |
+
| devign/vulnerability | 120.000 | 0.667 | 0.567 | 0.100 |
|
| 141 |
+
| emotion/6way | 120.000 | 0.492 | 0.367 | 0.125 |
|
| 142 |
+
| gsm8k/math | 120.000 | 0.733 | 0.125 | 0.608 |
|
| 143 |
+
| hellaswag/continuation | 120.000 | 0.608 | 0.292 | 0.317 |
|
| 144 |
+
| hotpotqa/comparison_yes_no | 26.000 | 0.923 | 0.462 | 0.462 |
|
| 145 |
+
| hotpotqa/retrieve | 120.000 | 0.875 | 0.292 | 0.583 |
|
| 146 |
+
| humaneval/completion | 119.000 | 0.849 | 0.513 | 0.336 |
|
| 147 |
+
| jailbreak/detect | 120.000 | 0.983 | 0.567 | 0.417 |
|
| 148 |
+
| long/contract_clause | 120.000 | 1.000 | 0.233 | 0.767 |
|
| 149 |
+
| long/email_thread | 120.000 | 1.000 | 0.050 | 0.950 |
|
| 150 |
+
| long/service_log | 120.000 | 1.000 | 0.242 | 0.758 |
|
| 151 |
+
| massive_en/intent | 120.000 | 0.950 | 0.742 | 0.208 |
|
| 152 |
+
| mbpp/bugspot | 120.000 | 0.858 | 0.575 | 0.283 |
|
| 153 |
+
| mbpp/solution | 120.000 | 1.000 | 0.850 | 0.150 |
|
| 154 |
+
| mmlu/mcq | 120.000 | 0.408 | 0.308 | 0.100 |
|
| 155 |
+
| mnli/claim | 120.000 | 0.875 | 0.317 | 0.558 |
|
| 156 |
+
| openbookqa/mcq | 120.000 | 0.650 | 0.283 | 0.367 |
|
| 157 |
+
| policy/access_control_transfer | 120.000 | 1.000 | 0.233 | 0.767 |
|
| 158 |
+
| policy/count_threshold_transfer | 120.000 | 0.867 | 0.108 | 0.758 |
|
| 159 |
+
| policy/free_shipping_transfer | 120.000 | 0.925 | 0.708 | 0.217 |
|
| 160 |
+
| policy/invoice_overdue_transfer | 120.000 | 0.808 | 0.658 | 0.150 |
|
| 161 |
+
| policy/invoice_total_transfer | 120.000 | 0.450 | 0.525 | -0.075 |
|
| 162 |
+
| policy/refund_approval_transfer | 120.000 | 0.958 | 0.325 | 0.633 |
|
| 163 |
+
| policy/return_window_transfer | 120.000 | 1.000 | 0.342 | 0.658 |
|
| 164 |
+
| policy/sla_urgency_transfer | 120.000 | 0.517 | 0.442 | 0.075 |
|
| 165 |
+
| policy/table_compare_transfer | 120.000 | 0.917 | 0.492 | 0.425 |
|
| 166 |
+
| policy/table_count_transfer | 120.000 | 0.317 | 0.158 | 0.158 |
|
| 167 |
+
| policy/table_extreme_transfer | 120.000 | 0.925 | 0.142 | 0.783 |
|
| 168 |
+
| prompt_injections/detect | 116.000 | 0.595 | 0.543 | 0.052 |
|
| 169 |
+
| qasc/mcq | 120.000 | 0.992 | 0.858 | 0.133 |
|
| 170 |
+
| sciq/mcq | 120.000 | 0.967 | 0.867 | 0.100 |
|
| 171 |
+
| scitail/support | 120.000 | 0.958 | 0.425 | 0.533 |
|
| 172 |
+
| snli/contradicts | 120.000 | 1.000 | 0.358 | 0.642 |
|
| 173 |
+
| snli/must_be_true | 120.000 | 0.992 | 0.925 | 0.067 |
|
| 174 |
+
| snli/nli | 120.000 | 0.908 | 0.342 | 0.567 |
|
| 175 |
+
| sst5/score | 120.000 | 0.383 | 0.283 | 0.100 |
|
| 176 |
+
| tev1_test/ag_news | 120.000 | 0.925 | 0.425 | 0.500 |
|
| 177 |
+
| tev1_test/banking77 | 120.000 | 0.717 | 0.550 | 0.167 |
|
| 178 |
+
| tev1_test/boolq | 120.000 | 0.842 | 0.558 | 0.283 |
|
| 179 |
+
| tev1_test/mnli | 120.000 | 0.750 | 0.392 | 0.358 |
|
| 180 |
+
| tev1_test/policy | 120.000 | 0.500 | 0.408 | 0.092 |
|
| 181 |
+
| tev1_test/routing | 120.000 | 0.442 | 0.292 | 0.150 |
|
| 182 |
+
| tev1_test/sst5 | 120.000 | 0.392 | 0.275 | 0.117 |
|
| 183 |
+
| triage/support_email | 120.000 | 0.958 | 0.392 | 0.567 |
|
| 184 |
+
| typed_decisions/agent_trace_observability | 120.000 | 0.808 | 0.333 | 0.475 |
|
| 185 |
+
| typed_decisions/customer_service | 120.000 | 0.742 | 0.283 | 0.458 |
|
| 186 |
+
| typed_decisions/invoice_processing | 120.000 | 0.817 | 0.475 | 0.342 |
|
| 187 |
+
| typed_decisions/security_incidents | 120.000 | 0.758 | 0.500 | 0.258 |
|
| 188 |
+
| winogrande/blank | 120.000 | 0.692 | 0.558 | 0.133 |
|
| 189 |
+
| yelp/score | 120.000 | 0.625 | 0.325 | 0.300 |
|
| 190 |
+
|
| 191 |
+
### Tev1 benchmark (tev1 v2.1 test records, held out)
|
| 192 |
+
|
| 193 |
+
| task | n | accuracy | source also in LightDec training |
|
| 194 |
+
|---|---|---|---|
|
| 195 |
+
| tev1_test/ag_news | 150 | 0.920 | yes |
|
| 196 |
+
| tev1_test/banking77 | 200 | 0.720 | no |
|
| 197 |
+
| tev1_test/boolq | 200 | 0.835 | yes |
|
| 198 |
+
| tev1_test/mnli | 300 | 0.723 | yes |
|
| 199 |
+
| tev1_test/policy | 1200 | 0.529 | no |
|
| 200 |
+
| tev1_test/routing | 600 | 0.458 | no |
|
| 201 |
+
| tev1_test/sst5 | 150 | 0.393 | no |
|
| 202 |
+
|
| 203 |
+
All categories 0.584 on 2800 decisions; excluding sources LightDec also trains on 0.518. Tev1-4B's published results (880/1,000 main decisions,
|
| 204 |
+
300/300 policy transfer) are reused development benchmarks and may not be the same records.
|
| 205 |
+
|
| 206 |
+
### Support-email triage (`triage/support_email`, 101 held-out emails)
|
| 207 |
+
|
| 208 |
+
Per question: topic 0.990 · refund 1.000 · breakage 0.990 · anger 0.921 · judgment 0.822. Routing-pile accuracy (person / engineering / feature log / reply): 0.941.
|
| 209 |
+
|
| 210 |
+
### Long-state tasks at max_len 2048
|
| 211 |
+
|
| 212 |
+
long/contract_clause 1.000 (chance 0.20) · long/email_thread 1.000 (chance 0.25) · long/service_log 1.000 (chance 0.20)
|
| 213 |
+
|
| 214 |
+
### Latency (measured on NVIDIA RTX PRO 6000 Blackwell Server Edition)
|
| 215 |
+
|
| 216 |
+
| device | weights | questions per call | p50 ms | p95 ms |
|
| 217 |
+
|---|---|---|---|---|
|
| 218 |
+
| cuda | fp16 | 1 | 10.1 | 10.4 |
|
| 219 |
+
| cuda | fp16 | 5 | 11.3 | 11.4 |
|
| 220 |
+
| cuda | fp16 | 10 | 12.6 | 12.7 |
|
| 221 |
+
| cpu fp32 | int8 file | 1 | 48.4 | 48.8 |
|
| 222 |
+
|
| 223 |
+
Batched throughput: 2589 decisions/s.
|
| 224 |
+
|
| 225 |
+
## Training
|
| 226 |
+
|
| 227 |
+
Strictly proper objective (log + 0.5·spherical + 1.0·RPS for ordinal), soft teacher targets where available,
|
| 228 |
+
task-balanced sampling (α = 0.5), NOTA augmentation (8%), per-epoch option reshuffling,
|
| 229 |
+
AdamW + LLRD 0.9, EMA 0.999, 10 epoch(s), RLCD stage off/reverted.
|
| 230 |
+
Training time 291.7 min on NVIDIA RTX PRO 6000 Blackwell Server Edition.
|
| 231 |
+
|
| 232 |
+
## Limitations
|
| 233 |
+
|
| 234 |
+
* Closed world: the model ranks the options you give it. Add "None of the above" when that is a valid answer (it was trained with it).
|
| 235 |
+
* Multi-step arithmetic (GSM8K / AQuA) is a System-2 task; route low-confidence answers to an LLM (`defer` flag, threshold 0.7).
|
| 236 |
+
* English only; code judgement (vulnerabilities, clones) is still weak; check calibration on your own traffic.
|
| 237 |
+
|
| 238 |
+
## References
|
| 239 |
+
|
| 240 |
+
Warner et al. 2024 (ModernBERT) · Weller et al. 2025 (Ettin) · Gneiting & Raftery 2007 (strictly proper scoring rules) ·
|
| 241 |
+
Guo et al. 2017 (temperature scaling) · Geifman & El-Yaniv 2017 (selective classification) · Shao et al. 2024 (GRPO) ·
|
| 242 |
+
Williams 1992 (REINFORCE) · Zaheer et al. 2017 / Lee et al. 2019 (Deep Sets / Set Transformer) · Hinton et al. 2015 (distillation) ·
|
| 243 |
+
Laya (convaiinnovations) · TypeSafe Jev · Together Tev1.
|
compact-int8/encoder/config.json
ADDED
|
@@ -0,0 +1,79 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"architectures": [
|
| 3 |
+
"ModernBertForMaskedLM"
|
| 4 |
+
],
|
| 5 |
+
"attention_bias": false,
|
| 6 |
+
"attention_dropout": 0.0,
|
| 7 |
+
"bos_token_id": 50281,
|
| 8 |
+
"causal_mask": false,
|
| 9 |
+
"classifier_activation": "gelu",
|
| 10 |
+
"classifier_bias": false,
|
| 11 |
+
"classifier_dropout": 0.0,
|
| 12 |
+
"classifier_pooling": "mean",
|
| 13 |
+
"cls_token_id": 50281,
|
| 14 |
+
"decoder_bias": true,
|
| 15 |
+
"deterministic_flash_attn": false,
|
| 16 |
+
"dtype": "float32",
|
| 17 |
+
"embedding_dropout": 0.0,
|
| 18 |
+
"eos_token_id": 50282,
|
| 19 |
+
"global_attn_every_n_layers": 3,
|
| 20 |
+
"gradient_checkpointing": false,
|
| 21 |
+
"hidden_activation": "gelu",
|
| 22 |
+
"hidden_size": 768,
|
| 23 |
+
"initializer_cutoff_factor": 2.0,
|
| 24 |
+
"initializer_range": 0.02,
|
| 25 |
+
"intermediate_size": 1152,
|
| 26 |
+
"is_causal": false,
|
| 27 |
+
"layer_norm_eps": 1e-05,
|
| 28 |
+
"layer_types": [
|
| 29 |
+
"full_attention",
|
| 30 |
+
"sliding_attention",
|
| 31 |
+
"sliding_attention",
|
| 32 |
+
"full_attention",
|
| 33 |
+
"sliding_attention",
|
| 34 |
+
"sliding_attention",
|
| 35 |
+
"full_attention",
|
| 36 |
+
"sliding_attention",
|
| 37 |
+
"sliding_attention",
|
| 38 |
+
"full_attention",
|
| 39 |
+
"sliding_attention",
|
| 40 |
+
"sliding_attention",
|
| 41 |
+
"full_attention",
|
| 42 |
+
"sliding_attention",
|
| 43 |
+
"sliding_attention",
|
| 44 |
+
"full_attention",
|
| 45 |
+
"sliding_attention",
|
| 46 |
+
"sliding_attention",
|
| 47 |
+
"full_attention",
|
| 48 |
+
"sliding_attention",
|
| 49 |
+
"sliding_attention",
|
| 50 |
+
"full_attention"
|
| 51 |
+
],
|
| 52 |
+
"local_attention": 128,
|
| 53 |
+
"max_position_embeddings": 7999,
|
| 54 |
+
"mlp_bias": false,
|
| 55 |
+
"mlp_dropout": 0.0,
|
| 56 |
+
"model_type": "modernbert",
|
| 57 |
+
"norm_bias": false,
|
| 58 |
+
"norm_eps": 1e-05,
|
| 59 |
+
"num_attention_heads": 12,
|
| 60 |
+
"num_hidden_layers": 22,
|
| 61 |
+
"pad_token_id": 50283,
|
| 62 |
+
"position_embedding_type": "sans_pos",
|
| 63 |
+
"rope_parameters": {
|
| 64 |
+
"full_attention": {
|
| 65 |
+
"rope_theta": 160000.0,
|
| 66 |
+
"rope_type": "default"
|
| 67 |
+
},
|
| 68 |
+
"sliding_attention": {
|
| 69 |
+
"rope_theta": 160000.0,
|
| 70 |
+
"rope_type": "default"
|
| 71 |
+
}
|
| 72 |
+
},
|
| 73 |
+
"sep_token_id": 50282,
|
| 74 |
+
"sparse_pred_ignore_index": -100,
|
| 75 |
+
"sparse_prediction": false,
|
| 76 |
+
"tie_word_embeddings": true,
|
| 77 |
+
"transformers_version": "5.17.0",
|
| 78 |
+
"vocab_size": 50368
|
| 79 |
+
}
|
compact-int8/falcondec_config.json
ADDED
|
@@ -0,0 +1,54 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"max_len": 2048,
|
| 3 |
+
"long_max_len": 2048,
|
| 4 |
+
"long_opts_threshold": 24,
|
| 5 |
+
"head_max_len": 192,
|
| 6 |
+
"max_tok_per_opt": 24,
|
| 7 |
+
"max_opts_single_pass": 96,
|
| 8 |
+
"interact_layers": 2,
|
| 9 |
+
"interact_heads": 8,
|
| 10 |
+
"name": "LightDec_V2_Long",
|
| 11 |
+
"backbone": "jhu-clsp/ettin-encoder-150m",
|
| 12 |
+
"backbone_init": "pretrained",
|
| 13 |
+
"lineage": [
|
| 14 |
+
"jhu-clsp/ettin-encoder-150m"
|
| 15 |
+
],
|
| 16 |
+
"parent_version": null,
|
| 17 |
+
"special": {
|
| 18 |
+
"cls": 50281,
|
| 19 |
+
"sep": 50282,
|
| 20 |
+
"mask": 50284,
|
| 21 |
+
"pad": 50283
|
| 22 |
+
},
|
| 23 |
+
"defer_threshold": 0.7,
|
| 24 |
+
"qtypes": {
|
| 25 |
+
"choice": 0,
|
| 26 |
+
"noul": 1,
|
| 27 |
+
"score": 2
|
| 28 |
+
},
|
| 29 |
+
"n_buckets": 4,
|
| 30 |
+
"version": "1.0.0",
|
| 31 |
+
"notebook_version": "V3.7",
|
| 32 |
+
"created": "2026-09-28T10:52:36.426087Z",
|
| 33 |
+
"temperature": [
|
| 34 |
+
[
|
| 35 |
+
2.37591814994812,
|
| 36 |
+
2.0426506996154785,
|
| 37 |
+
1.6455482244491577,
|
| 38 |
+
1.8417285680770874
|
| 39 |
+
],
|
| 40 |
+
[
|
| 41 |
+
2.4885058403015137,
|
| 42 |
+
2.4885058403015137,
|
| 43 |
+
2.4885058403015137,
|
| 44 |
+
2.4885058403015137
|
| 45 |
+
],
|
| 46 |
+
[
|
| 47 |
+
2.2925639152526855,
|
| 48 |
+
2.2925639152526855,
|
| 49 |
+
2.2925639152526855,
|
| 50 |
+
2.2925639152526855
|
| 51 |
+
]
|
| 52 |
+
],
|
| 53 |
+
"weights": "model_int8.safetensors"
|
| 54 |
+
}
|
compact-int8/falcondec_modeling.py
ADDED
|
@@ -0,0 +1,358 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# -*- coding: utf-8 -*-
|
| 2 |
+
"""FalconDec — Falcon Decision model: single-pass, typed, calibrated closed-set decisions.
|
| 3 |
+
|
| 4 |
+
Layout : [CLS] question [SEP] [MASK] opt_1 ... [MASK] opt_k [SEP] state [SEP]
|
| 5 |
+
Head : marker vectors + CLS context + question-type embedding
|
| 6 |
+
-> permutation-equivariant set transformer (options attend to each other; no positions)
|
| 7 |
+
-> MLP -> one logit per option -> softmax over this question's options
|
| 8 |
+
Calib. : temperature per (question type, option-count bucket), stored in the checkpoint
|
| 9 |
+
"""
|
| 10 |
+
from __future__ import annotations
|
| 11 |
+
|
| 12 |
+
import contextlib
|
| 13 |
+
import json
|
| 14 |
+
import shutil
|
| 15 |
+
from pathlib import Path
|
| 16 |
+
|
| 17 |
+
import numpy as np
|
| 18 |
+
import torch
|
| 19 |
+
import torch.nn as nn
|
| 20 |
+
|
| 21 |
+
QTYPES = {"choice": 0, "noul": 1, "score": 2}
|
| 22 |
+
N_BUCKETS = 4
|
| 23 |
+
CONFIG_FILE = "falcondec_config.json"
|
| 24 |
+
FP16_FILE = "model.safetensors"
|
| 25 |
+
INT8_FILE = "model_int8.safetensors"
|
| 26 |
+
MODEL_KEYS = ("input_ids", "attention_mask", "marker_pos", "marker_mask", "qtype")
|
| 27 |
+
|
| 28 |
+
|
| 29 |
+
def n_bucket(n: int) -> int:
|
| 30 |
+
return 0 if n <= 2 else 1 if n <= 5 else 2 if n <= 12 else 3
|
| 31 |
+
|
| 32 |
+
|
| 33 |
+
def special_ids(tok) -> dict:
|
| 34 |
+
sp = {"cls": tok.cls_token_id, "sep": tok.sep_token_id, "mask": tok.mask_token_id, "pad": tok.pad_token_id}
|
| 35 |
+
if sp["cls"] is None:
|
| 36 |
+
sp["cls"] = tok.bos_token_id
|
| 37 |
+
if sp["sep"] is None:
|
| 38 |
+
sp["sep"] = tok.eos_token_id
|
| 39 |
+
if sp["pad"] is None:
|
| 40 |
+
sp["pad"] = sp["sep"]
|
| 41 |
+
if sp["mask"] is None:
|
| 42 |
+
raise ValueError("FalconDec needs a tokenizer with a mask token (used as the option marker).")
|
| 43 |
+
return {k: int(v) for k, v in sp.items()}
|
| 44 |
+
|
| 45 |
+
|
| 46 |
+
def _as_list(x):
|
| 47 |
+
return x.tolist() if hasattr(x, "tolist") else list(x)
|
| 48 |
+
|
| 49 |
+
|
| 50 |
+
def assemble(q_ids, opt_ids, s_ids, sp, max_len=512, head_max_len=192, max_tok_per_opt=24,
|
| 51 |
+
long_max_len=2048, long_opts_threshold=24):
|
| 52 |
+
"""Build one input sequence. Returns (ids, marker_positions)."""
|
| 53 |
+
n = len(opt_ids)
|
| 54 |
+
eff = max_len if n <= long_opts_threshold else max(max_len, long_max_len)
|
| 55 |
+
q = _as_list(q_ids)[:96]
|
| 56 |
+
need = len(q) + 2 + n * (max_tok_per_opt + 1)
|
| 57 |
+
head = min(eff - 64, max(head_max_len, need))
|
| 58 |
+
per = max(2, min(max_tok_per_opt, (head - len(q) - 2) // max(n, 1) - 1))
|
| 59 |
+
ids = [sp["cls"]] + q + [sp["sep"]]
|
| 60 |
+
markers = []
|
| 61 |
+
for o in opt_ids:
|
| 62 |
+
markers.append(len(ids))
|
| 63 |
+
ids.append(sp["mask"])
|
| 64 |
+
ids.extend(_as_list(o[:per]))
|
| 65 |
+
ids.append(sp["sep"])
|
| 66 |
+
room = eff - len(ids) - 1
|
| 67 |
+
if room > 0 and s_ids is not None and len(s_ids) > 0:
|
| 68 |
+
ids.extend(_as_list(s_ids[:room]))
|
| 69 |
+
ids.append(sp["sep"])
|
| 70 |
+
if len(ids) > eff:
|
| 71 |
+
ids = ids[:eff]
|
| 72 |
+
if markers and markers[-1] >= len(ids):
|
| 73 |
+
raise ValueError("Too many options for one pass; reduce options or raise long_max_len.")
|
| 74 |
+
return ids, markers
|
| 75 |
+
|
| 76 |
+
|
| 77 |
+
def collate_features(feats, pad_id, device=None):
|
| 78 |
+
"""feats: list of (ids, markers, qtype_index) -> dict of padded tensors matching FalconDec.forward."""
|
| 79 |
+
B = len(feats)
|
| 80 |
+
T = max(len(f[0]) for f in feats)
|
| 81 |
+
K = max(len(f[1]) for f in feats)
|
| 82 |
+
ids = torch.full((B, T), pad_id, dtype=torch.long)
|
| 83 |
+
att = torch.zeros((B, T), dtype=torch.long)
|
| 84 |
+
mpos = torch.zeros((B, K), dtype=torch.long)
|
| 85 |
+
mmask = torch.zeros((B, K), dtype=torch.bool)
|
| 86 |
+
qt = torch.zeros(B, dtype=torch.long)
|
| 87 |
+
for i, (x, mk, q) in enumerate(feats):
|
| 88 |
+
ids[i, : len(x)] = torch.as_tensor(x, dtype=torch.long)
|
| 89 |
+
att[i, : len(x)] = 1
|
| 90 |
+
mpos[i, : len(mk)] = torch.as_tensor(mk, dtype=torch.long)
|
| 91 |
+
mmask[i, : len(mk)] = True
|
| 92 |
+
qt[i] = int(q)
|
| 93 |
+
out = dict(input_ids=ids, attention_mask=att, marker_pos=mpos, marker_mask=mmask, qtype=qt)
|
| 94 |
+
if device is not None:
|
| 95 |
+
out = {k: v.to(device, non_blocking=True) for k, v in out.items()}
|
| 96 |
+
return out
|
| 97 |
+
|
| 98 |
+
|
| 99 |
+
class OptionInteraction(nn.Module):
|
| 100 |
+
"""Set transformer over the options of one question (no positional encoding => order-equivariant)."""
|
| 101 |
+
|
| 102 |
+
def __init__(self, d, n_layers=2, n_heads=8, dropout=0.1):
|
| 103 |
+
super().__init__()
|
| 104 |
+
layer = nn.TransformerEncoderLayer(d, n_heads, dim_feedforward=2 * d, dropout=dropout,
|
| 105 |
+
activation="gelu", batch_first=True, norm_first=True)
|
| 106 |
+
self.enc = nn.TransformerEncoder(layer, n_layers, enable_nested_tensor=False)
|
| 107 |
+
|
| 108 |
+
def forward(self, x, mask):
|
| 109 |
+
return self.enc(x, src_key_padding_mask=~mask)
|
| 110 |
+
|
| 111 |
+
|
| 112 |
+
class FalconDec(nn.Module):
|
| 113 |
+
def __init__(self, encoder, fcfg: dict):
|
| 114 |
+
super().__init__()
|
| 115 |
+
self.encoder = encoder
|
| 116 |
+
self.fcfg = dict(fcfg)
|
| 117 |
+
d = encoder.config.hidden_size
|
| 118 |
+
self.qtype_emb = nn.Embedding(len(QTYPES), d)
|
| 119 |
+
self.ctx_proj = nn.Linear(d, d)
|
| 120 |
+
self.opt_norm = nn.LayerNorm(d)
|
| 121 |
+
self.interact = OptionInteraction(d, int(self.fcfg.get("interact_layers", 2)),
|
| 122 |
+
int(self.fcfg.get("interact_heads", 8)))
|
| 123 |
+
self.scorer = nn.Sequential(nn.Linear(d, d), nn.GELU(), nn.Dropout(0.1), nn.Linear(d, 1))
|
| 124 |
+
self.register_buffer("temperature", torch.ones(len(QTYPES), N_BUCKETS))
|
| 125 |
+
|
| 126 |
+
@property
|
| 127 |
+
def device(self):
|
| 128 |
+
return next(self.parameters()).device
|
| 129 |
+
|
| 130 |
+
def num_parameters(self):
|
| 131 |
+
return sum(p.numel() for p in self.parameters())
|
| 132 |
+
|
| 133 |
+
def forward(self, input_ids, attention_mask, marker_pos, marker_mask, qtype):
|
| 134 |
+
h = self.encoder(input_ids=input_ids, attention_mask=attention_mask).last_hidden_state
|
| 135 |
+
B, K = marker_pos.shape
|
| 136 |
+
idx = marker_pos.unsqueeze(-1).expand(B, K, h.size(-1))
|
| 137 |
+
opt = h.gather(1, idx)
|
| 138 |
+
ctx = self.ctx_proj(h[:, 0]).unsqueeze(1)
|
| 139 |
+
x = self.opt_norm(opt + ctx + self.qtype_emb(qtype).unsqueeze(1))
|
| 140 |
+
x = self.interact(x, marker_mask)
|
| 141 |
+
logits = self.scorer(x).squeeze(-1).float()
|
| 142 |
+
return logits.masked_fill(~marker_mask, -1e4)
|
| 143 |
+
|
| 144 |
+
|
| 145 |
+
# ------------------------------------------------------------------ storage
|
| 146 |
+
def clean_state_dict(sd):
|
| 147 |
+
return {k.replace("_orig_mod.", ""): v.detach().cpu().contiguous() for k, v in sd.items()}
|
| 148 |
+
|
| 149 |
+
|
| 150 |
+
def quantize_int8(sd, min_numel=4096):
|
| 151 |
+
"""Per-output-channel symmetric int8 for every matrix; fp16 for everything else."""
|
| 152 |
+
out = {}
|
| 153 |
+
for k, v in sd.items():
|
| 154 |
+
if v.is_floating_point() and v.ndim == 2 and v.numel() >= min_numel:
|
| 155 |
+
w = v.float()
|
| 156 |
+
s = (w.abs().amax(dim=1, keepdim=True) / 127.0).clamp_min(1e-12)
|
| 157 |
+
out[k] = torch.round(w / s).clamp_(-127, 127).to(torch.int8).contiguous()
|
| 158 |
+
out[k + "::scale"] = s.contiguous()
|
| 159 |
+
elif v.is_floating_point():
|
| 160 |
+
out[k] = v.half().contiguous()
|
| 161 |
+
else:
|
| 162 |
+
out[k] = v.contiguous()
|
| 163 |
+
return out
|
| 164 |
+
|
| 165 |
+
|
| 166 |
+
def dequantize_int8(sd):
|
| 167 |
+
out = {}
|
| 168 |
+
for k, v in sd.items():
|
| 169 |
+
if k.endswith("::scale"):
|
| 170 |
+
continue
|
| 171 |
+
s = sd.get(k + "::scale")
|
| 172 |
+
out[k] = (v.float() * s).half() if s is not None else v
|
| 173 |
+
return out
|
| 174 |
+
|
| 175 |
+
|
| 176 |
+
def save_falcondec(model, tok, out_dir, int8=False, extra_files=None):
|
| 177 |
+
from safetensors.torch import save_file
|
| 178 |
+
out = Path(out_dir)
|
| 179 |
+
out.mkdir(parents=True, exist_ok=True)
|
| 180 |
+
enc_cfg = model.encoder.config
|
| 181 |
+
if hasattr(enc_cfg, "reference_compile"):
|
| 182 |
+
enc_cfg.reference_compile = False
|
| 183 |
+
enc_cfg.save_pretrained(str(out / "encoder"))
|
| 184 |
+
tok.save_pretrained(str(out / "tokenizer"))
|
| 185 |
+
sd = clean_state_dict(model.state_dict())
|
| 186 |
+
fc = dict(model.fcfg)
|
| 187 |
+
fc["temperature"] = model.temperature.detach().float().cpu().tolist()
|
| 188 |
+
fc["weights"] = INT8_FILE if int8 else FP16_FILE
|
| 189 |
+
if int8:
|
| 190 |
+
save_file(quantize_int8(sd), str(out / INT8_FILE), metadata={"format": "falcondec-int8"})
|
| 191 |
+
else:
|
| 192 |
+
save_file({k: (v.half() if v.is_floating_point() else v) for k, v in sd.items()},
|
| 193 |
+
str(out / FP16_FILE), metadata={"format": "falcondec-fp16"})
|
| 194 |
+
(out / CONFIG_FILE).write_text(json.dumps(fc, indent=2, default=str), encoding="utf-8")
|
| 195 |
+
try:
|
| 196 |
+
shutil.copy(__file__, out / "falcondec_modeling.py")
|
| 197 |
+
except Exception:
|
| 198 |
+
pass
|
| 199 |
+
for name, content in (extra_files or {}).items():
|
| 200 |
+
(out / name).write_text(content, encoding="utf-8")
|
| 201 |
+
return out
|
| 202 |
+
|
| 203 |
+
|
| 204 |
+
def load_falcondec(path, device=None, dtype=None, attn_implementation="sdpa"):
|
| 205 |
+
"""Load a FalconDec directory (fp16 or int8) or Hub repo. Returns (model, tokenizer)."""
|
| 206 |
+
from safetensors.torch import load_file
|
| 207 |
+
from transformers import AutoConfig, AutoModel, AutoTokenizer
|
| 208 |
+
p = Path(path)
|
| 209 |
+
if not p.exists():
|
| 210 |
+
from huggingface_hub import snapshot_download
|
| 211 |
+
p = Path(snapshot_download(str(path)))
|
| 212 |
+
fc = json.loads((p / CONFIG_FILE).read_text(encoding="utf-8"))
|
| 213 |
+
ecfg = AutoConfig.from_pretrained(str(p / "encoder"))
|
| 214 |
+
if hasattr(ecfg, "reference_compile"):
|
| 215 |
+
ecfg.reference_compile = False
|
| 216 |
+
try:
|
| 217 |
+
enc = AutoModel.from_config(ecfg, attn_implementation=attn_implementation)
|
| 218 |
+
except Exception:
|
| 219 |
+
enc = AutoModel.from_config(ecfg)
|
| 220 |
+
model = FalconDec(enc, fc)
|
| 221 |
+
wf = p / fc.get("weights", FP16_FILE)
|
| 222 |
+
if not wf.exists():
|
| 223 |
+
wf = p / (INT8_FILE if (p / INT8_FILE).exists() else FP16_FILE)
|
| 224 |
+
sd = load_file(str(wf))
|
| 225 |
+
if wf.name == INT8_FILE:
|
| 226 |
+
sd = dequantize_int8(sd)
|
| 227 |
+
missing, unexpected = model.load_state_dict(sd, strict=False)
|
| 228 |
+
if missing or unexpected:
|
| 229 |
+
print(f"[FalconDec] load warning: missing={list(missing)[:5]} unexpected={list(unexpected)[:5]}")
|
| 230 |
+
tok = AutoTokenizer.from_pretrained(str(p / "tokenizer"))
|
| 231 |
+
dev = torch.device(device) if device is not None else torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
| 232 |
+
model.to(dev)
|
| 233 |
+
if dtype is not None:
|
| 234 |
+
model.to(dtype)
|
| 235 |
+
model.eval()
|
| 236 |
+
return model, tok
|
| 237 |
+
|
| 238 |
+
|
| 239 |
+
# ------------------------------------------------------------------ inference
|
| 240 |
+
def _amp(model):
|
| 241 |
+
if model.device.type == "cuda" and next(model.parameters()).dtype == torch.float32:
|
| 242 |
+
dt = torch.bfloat16 if torch.cuda.is_bf16_supported() else torch.float16
|
| 243 |
+
return torch.autocast("cuda", dtype=dt)
|
| 244 |
+
return contextlib.nullcontext()
|
| 245 |
+
|
| 246 |
+
|
| 247 |
+
def _state_text(state):
|
| 248 |
+
if state is None:
|
| 249 |
+
return ""
|
| 250 |
+
return state if isinstance(state, str) else json.dumps(state, ensure_ascii=False)
|
| 251 |
+
|
| 252 |
+
|
| 253 |
+
@torch.no_grad()
|
| 254 |
+
def score_items(model, tok, items, batch_size=32):
|
| 255 |
+
"""items: [{"state", "question", "options", "type"?, "option_tokens"?, "seq_len"?}] -> list of prob arrays."""
|
| 256 |
+
fc = model.fcfg
|
| 257 |
+
M = int(fc.get("max_opts_single_pass", 96))
|
| 258 |
+
out = [None] * len(items)
|
| 259 |
+
small = [i for i, it in enumerate(items) if len(it["options"]) <= M]
|
| 260 |
+
for i, it in enumerate(items):
|
| 261 |
+
if len(it["options"]) > M:
|
| 262 |
+
out[i] = _score_large(model, tok, it, batch_size)
|
| 263 |
+
if not small:
|
| 264 |
+
return out
|
| 265 |
+
sp = fc["special"]
|
| 266 |
+
prepared = {}
|
| 267 |
+
for i in small:
|
| 268 |
+
it = items[i]
|
| 269 |
+
if len(it["options"]) < 2:
|
| 270 |
+
raise ValueError("Every question needs at least two options.")
|
| 271 |
+
budget = int(it.get("option_tokens", fc["max_tok_per_opt"]))
|
| 272 |
+
q_ids = tok(it.get("question", ""), add_special_tokens=False, truncation=True, max_length=96)["input_ids"]
|
| 273 |
+
o_ids = tok([str(o) for o in it["options"]], add_special_tokens=False, truncation=True,
|
| 274 |
+
max_length=budget)["input_ids"]
|
| 275 |
+
s = _state_text(it.get("state"))
|
| 276 |
+
s_ids = tok(s, add_special_tokens=False, truncation=True,
|
| 277 |
+
max_length=int(fc["long_max_len"]))["input_ids"] if s else []
|
| 278 |
+
ids, mk = assemble(q_ids, o_ids, s_ids, sp, int(it.get("seq_len", fc["max_len"])), fc["head_max_len"],
|
| 279 |
+
budget, fc["long_max_len"], fc["long_opts_threshold"])
|
| 280 |
+
prepared[i] = (ids, mk, QTYPES[it.get("type", "choice")])
|
| 281 |
+
order = sorted(small, key=lambda i: len(prepared[i][0]))
|
| 282 |
+
model.eval()
|
| 283 |
+
for b0 in range(0, len(order), batch_size):
|
| 284 |
+
idxs = order[b0: b0 + batch_size]
|
| 285 |
+
batch = collate_features([prepared[i] for i in idxs], sp["pad"], model.device)
|
| 286 |
+
with _amp(model):
|
| 287 |
+
logits = model(**batch)
|
| 288 |
+
for j, i in enumerate(idxs):
|
| 289 |
+
n = len(prepared[i][1])
|
| 290 |
+
T = model.temperature[prepared[i][2], n_bucket(n)].float().clamp_min(1e-3)
|
| 291 |
+
out[i] = torch.softmax(logits[j, :n].float() / T, -1).cpu().numpy()
|
| 292 |
+
return out
|
| 293 |
+
|
| 294 |
+
|
| 295 |
+
def _score_large(model, tok, it, batch_size):
|
| 296 |
+
"""Tournament for > max_opts_single_pass options: keep the best of each chunk, then one final pass."""
|
| 297 |
+
M = int(model.fcfg.get("max_opts_single_pass", 96))
|
| 298 |
+
opts = list(it["options"])
|
| 299 |
+
cand = list(range(len(opts)))
|
| 300 |
+
while len(cand) > M:
|
| 301 |
+
groups = [cand[i: i + M] for i in range(0, len(cand), M)]
|
| 302 |
+
keep = max(1, M // len(groups))
|
| 303 |
+
ps = score_items(model, tok, [dict(it, options=[opts[c] for c in g]) for g in groups], batch_size)
|
| 304 |
+
cand = [g[t] for g, p in zip(groups, ps) for t in np.argsort(-p)[:keep]]
|
| 305 |
+
p = score_items(model, tok, [dict(it, options=[opts[c] for c in cand])], batch_size)[0]
|
| 306 |
+
full = np.zeros(len(opts), dtype=np.float32)
|
| 307 |
+
full[cand] = p
|
| 308 |
+
return full
|
| 309 |
+
|
| 310 |
+
|
| 311 |
+
def _normalize_question(q):
|
| 312 |
+
qtype = q.get("type", "choice")
|
| 313 |
+
text = q.get("question") or q.get("instructions") or ""
|
| 314 |
+
if qtype == "noul":
|
| 315 |
+
lab = q.get("labels") or {}
|
| 316 |
+
return qtype, text, [True, False], [str(lab.get("true", "Yes")), str(lab.get("false", "No"))]
|
| 317 |
+
crit = q.get("criteria", q.get("options"))
|
| 318 |
+
if qtype == "score":
|
| 319 |
+
opts = [str(c) for c in crit]
|
| 320 |
+
return qtype, text, list(range(len(opts))), opts
|
| 321 |
+
if isinstance(crit, dict):
|
| 322 |
+
return qtype, text, list(crit), [f"{k}: {v}" if v else str(k) for k, v in crit.items()]
|
| 323 |
+
return qtype, text, list(crit), [str(c) for c in crit]
|
| 324 |
+
|
| 325 |
+
|
| 326 |
+
@torch.no_grad()
|
| 327 |
+
def decide(model, tok, state, questions, defer_threshold=None, batch_size=32):
|
| 328 |
+
"""Answer typed questions about one state.
|
| 329 |
+
|
| 330 |
+
questions: list of dicts, or a Jev/Laya-style dict {key: question}. Each question:
|
| 331 |
+
{"type": "choice"|"noul"|"score", "question"/"instructions": str,
|
| 332 |
+
"options": [...] or "criteria": {key: description} / [levels], "option_tokens"?: int}
|
| 333 |
+
"""
|
| 334 |
+
if isinstance(questions, dict):
|
| 335 |
+
questions = [dict(q, key=k) for k, q in questions.items()]
|
| 336 |
+
items, metas = [], []
|
| 337 |
+
for q in questions:
|
| 338 |
+
qtype, text, keys, opts = _normalize_question(q)
|
| 339 |
+
item = {"state": state, "question": text, "options": opts, "type": qtype}
|
| 340 |
+
for extra in ("option_tokens", "seq_len"):
|
| 341 |
+
if extra in q:
|
| 342 |
+
item[extra] = q[extra]
|
| 343 |
+
items.append(item)
|
| 344 |
+
metas.append((q, qtype, keys, opts))
|
| 345 |
+
probs = score_items(model, tok, items, batch_size)
|
| 346 |
+
thr = model.fcfg.get("defer_threshold", 0.0) if defer_threshold is None else defer_threshold
|
| 347 |
+
results = []
|
| 348 |
+
for (q, qtype, keys, opts), p in zip(metas, probs):
|
| 349 |
+
i = int(np.argmax(p))
|
| 350 |
+
r = {"key": q.get("key"), "type": qtype, "question": items[len(results)]["question"],
|
| 351 |
+
"choice": keys[i], "choice_text": opts[i], "confidence": float(p[i]),
|
| 352 |
+
"probs": {str(k): float(v) for k, v in zip(keys, p)}, "defer": float(p[i]) < thr}
|
| 353 |
+
if qtype == "noul":
|
| 354 |
+
r["p_true"] = float(p[0])
|
| 355 |
+
if qtype == "score":
|
| 356 |
+
r["expected_level"] = float(np.dot(p, np.arange(len(p))))
|
| 357 |
+
results.append(r)
|
| 358 |
+
return {"results": results, "answers": {r["key"]: r for r in results if r["key"] is not None}}
|
compact-int8/falcondec_report.json
ADDED
|
@@ -0,0 +1,1802 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"version": "1.0.0",
|
| 3 |
+
"notebook_version": "V3.7",
|
| 4 |
+
"mode": "scratch",
|
| 5 |
+
"parent_version": null,
|
| 6 |
+
"lineage": [
|
| 7 |
+
"jhu-clsp/ettin-encoder-150m"
|
| 8 |
+
],
|
| 9 |
+
"config": {
|
| 10 |
+
"MODE": "scratch",
|
| 11 |
+
"FINETUNE_FROM": "latest",
|
| 12 |
+
"MODEL_NAME": "LightDec_V2_Long",
|
| 13 |
+
"BACKBONE": "jhu-clsp/ettin-encoder-150m",
|
| 14 |
+
"BACKBONE_INIT": "pretrained",
|
| 15 |
+
"PRESET": "long",
|
| 16 |
+
"CUSTOM_DATA_JSONL": "",
|
| 17 |
+
"CUSTOM_WEIGHT": 3.0,
|
| 18 |
+
"INCLUDE_BUILDERS": [],
|
| 19 |
+
"EXCLUDE_BUILDERS": [
|
| 20 |
+
"mind2web"
|
| 21 |
+
],
|
| 22 |
+
"TEV1_DIR": "./tev1",
|
| 23 |
+
"TEV1_AUTOBUILD": true,
|
| 24 |
+
"TEV1_POLICY_CAP": 13500,
|
| 25 |
+
"TEV1_ROUTING_CAP": 6000,
|
| 26 |
+
"TEV1_RESEARCH_CAP": 1000,
|
| 27 |
+
"TRIAGE_EMAILS": 200,
|
| 28 |
+
"TEV1_RETRIES": 3,
|
| 29 |
+
"TEV1_STEP_TIMEOUT_MIN": 180,
|
| 30 |
+
"TEV1_STALL_MIN": 15,
|
| 31 |
+
"RETRY_FAILED_BUILDERS": true,
|
| 32 |
+
"LONG_TASK_TRAIN": 1500,
|
| 33 |
+
"MAX_LEN": 2048,
|
| 34 |
+
"LONG_MAX_LEN": 2048,
|
| 35 |
+
"LONG_OPTS_THRESHOLD": 24,
|
| 36 |
+
"HEAD_MAX_LEN": 192,
|
| 37 |
+
"MAX_TOK_PER_OPT": 24,
|
| 38 |
+
"MAX_OPTS_SINGLE_PASS": 96,
|
| 39 |
+
"INTERACT_LAYERS": 2,
|
| 40 |
+
"INTERACT_HEADS": 8,
|
| 41 |
+
"BATCH_SIZE": 128,
|
| 42 |
+
"TOKENS_PER_BATCH": 65536,
|
| 43 |
+
"GRAD_ACCUM": 1,
|
| 44 |
+
"LR_ENCODER": 8e-05,
|
| 45 |
+
"LR_HEAD": 0.0006,
|
| 46 |
+
"LLRD": 0.9,
|
| 47 |
+
"WEIGHT_DECAY": 0.01,
|
| 48 |
+
"WARMUP_FRAC": 0.06,
|
| 49 |
+
"EPOCHS": 0,
|
| 50 |
+
"EMA_DECAY": 0.999,
|
| 51 |
+
"GRAD_CLIP": 1.0,
|
| 52 |
+
"FINETUNE_LR_SCALE": 0.5,
|
| 53 |
+
"SPHERICAL_W": 0.5,
|
| 54 |
+
"BRIER_W": 0.0,
|
| 55 |
+
"RPS_W": 1.0,
|
| 56 |
+
"TASK_SAMPLING_ALPHA": 0.5,
|
| 57 |
+
"NOTA_PROB": 0.08,
|
| 58 |
+
"RLCD_EPOCHS": 0.0,
|
| 59 |
+
"RLCD_SAMPLES": 4,
|
| 60 |
+
"RLCD_SIGMA": 0.3,
|
| 61 |
+
"RLCD_CE_ANCHOR": 0.5,
|
| 62 |
+
"DEVICE": "cuda",
|
| 63 |
+
"SEED": 42,
|
| 64 |
+
"NUM_WORKERS": 8,
|
| 65 |
+
"USE_COMPILE": true,
|
| 66 |
+
"ATTN_IMPL": "auto",
|
| 67 |
+
"OUTPUT_ROOT": "./falcondec_runs",
|
| 68 |
+
"EXPORT_INT8": true,
|
| 69 |
+
"SIZE_TARGET_MB": 250,
|
| 70 |
+
"DEFER_THRESHOLD": 0.7,
|
| 71 |
+
"EVAL_BASELINE_PROOF_V2": true,
|
| 72 |
+
"BASELINE_MAX_PER_TASK": 120,
|
| 73 |
+
"RUN_CPU_BENCH": true,
|
| 74 |
+
"PUSH_TO_HUB": false,
|
| 75 |
+
"HUB_REPO": ""
|
| 76 |
+
},
|
| 77 |
+
"epochs": 10,
|
| 78 |
+
"best_epoch": 4,
|
| 79 |
+
"rlcd_kept": false,
|
| 80 |
+
"history": [
|
| 81 |
+
{
|
| 82 |
+
"epoch": 1,
|
| 83 |
+
"train_loss": 0.7058714814313827,
|
| 84 |
+
"train_acc": 0.7342094450703411,
|
| 85 |
+
"val_macro": 0.8188568656508199,
|
| 86 |
+
"val_micro": 0.820702479338843,
|
| 87 |
+
"val_nll": 0.4076740937270036
|
| 88 |
+
},
|
| 89 |
+
{
|
| 90 |
+
"epoch": 2,
|
| 91 |
+
"train_loss": 0.431469622218984,
|
| 92 |
+
"train_acc": 0.8473141073247583,
|
| 93 |
+
"val_macro": 0.8450872188411825,
|
| 94 |
+
"val_micro": 0.8444628099173553,
|
| 95 |
+
"val_nll": 0.3623910443100994
|
| 96 |
+
},
|
| 97 |
+
{
|
| 98 |
+
"epoch": 3,
|
| 99 |
+
"train_loss": 0.34276796075090776,
|
| 100 |
+
"train_acc": 0.8791940521680108,
|
| 101 |
+
"val_macro": 0.8501904675940947,
|
| 102 |
+
"val_micro": 0.8499173553719008,
|
| 103 |
+
"val_nll": 0.39227130460368825
|
| 104 |
+
},
|
| 105 |
+
{
|
| 106 |
+
"epoch": 4,
|
| 107 |
+
"train_loss": 0.2773282952523419,
|
| 108 |
+
"train_acc": 0.9039863196679513,
|
| 109 |
+
"val_macro": 0.8519172452047367,
|
| 110 |
+
"val_micro": 0.8519834710743802,
|
| 111 |
+
"val_nll": 0.4529550437184426
|
| 112 |
+
},
|
| 113 |
+
{
|
| 114 |
+
"epoch": 5,
|
| 115 |
+
"train_loss": 0.22030453415052323,
|
| 116 |
+
"train_acc": 0.925851485588573,
|
| 117 |
+
"val_macro": 0.8488092001804082,
|
| 118 |
+
"val_micro": 0.85,
|
| 119 |
+
"val_nll": 0.5369876145373608
|
| 120 |
+
},
|
| 121 |
+
{
|
| 122 |
+
"epoch": 6,
|
| 123 |
+
"train_loss": 0.16953414903032543,
|
| 124 |
+
"train_acc": 0.9449156076226879,
|
| 125 |
+
"val_macro": 0.8474279008050457,
|
| 126 |
+
"val_micro": 0.8488842975206612,
|
| 127 |
+
"val_nll": 0.6705887193589142
|
| 128 |
+
},
|
| 129 |
+
{
|
| 130 |
+
"epoch": 7,
|
| 131 |
+
"train_loss": 0.126976966611751,
|
| 132 |
+
"train_acc": 0.9602833527848939,
|
| 133 |
+
"val_macro": 0.8453322694829587,
|
| 134 |
+
"val_micro": 0.8476446280991735,
|
| 135 |
+
"val_nll": 0.8430423012207805
|
| 136 |
+
},
|
| 137 |
+
{
|
| 138 |
+
"epoch": 8,
|
| 139 |
+
"train_loss": 0.09470793687766035,
|
| 140 |
+
"train_acc": 0.9714013975720073,
|
| 141 |
+
"val_macro": 0.8451769009269958,
|
| 142 |
+
"val_micro": 0.8467355371900827,
|
| 143 |
+
"val_nll": 1.0336680755654943
|
| 144 |
+
},
|
| 145 |
+
{
|
| 146 |
+
"epoch": 9,
|
| 147 |
+
"train_loss": 0.0710737590306488,
|
| 148 |
+
"train_acc": 0.9794225926173407,
|
| 149 |
+
"val_macro": 0.8435159403610613,
|
| 150 |
+
"val_micro": 0.8458677685950413,
|
| 151 |
+
"val_nll": 1.210751206553627
|
| 152 |
+
},
|
| 153 |
+
{
|
| 154 |
+
"epoch": 10,
|
| 155 |
+
"train_loss": 0.05665995772408678,
|
| 156 |
+
"train_acc": 0.9843832093140154,
|
| 157 |
+
"val_macro": 0.844272704258198,
|
| 158 |
+
"val_micro": 0.8474380165289256,
|
| 159 |
+
"val_nll": 1.41872131482475
|
| 160 |
+
}
|
| 161 |
+
],
|
| 162 |
+
"train_minutes": 291.67019440333047,
|
| 163 |
+
"temperatures": [
|
| 164 |
+
[
|
| 165 |
+
2.37591814994812,
|
| 166 |
+
2.0426506996154785,
|
| 167 |
+
1.6455482244491577,
|
| 168 |
+
1.8417285680770874
|
| 169 |
+
],
|
| 170 |
+
[
|
| 171 |
+
2.4885058403015137,
|
| 172 |
+
2.4885058403015137,
|
| 173 |
+
2.4885058403015137,
|
| 174 |
+
2.4885058403015137
|
| 175 |
+
],
|
| 176 |
+
[
|
| 177 |
+
2.2925639152526855,
|
| 178 |
+
2.2925639152526855,
|
| 179 |
+
2.2925639152526855,
|
| 180 |
+
2.2925639152526855
|
| 181 |
+
]
|
| 182 |
+
],
|
| 183 |
+
"data_counts": {
|
| 184 |
+
"train": 1023814,
|
| 185 |
+
"validation": 24200,
|
| 186 |
+
"test": 31990
|
| 187 |
+
},
|
| 188 |
+
"skipped_builders": [],
|
| 189 |
+
"test_overall": {
|
| 190 |
+
"n": 31990.0,
|
| 191 |
+
"micro_acc": 0.784057517974367,
|
| 192 |
+
"macro_acc": 0.7887006006875437,
|
| 193 |
+
"nll": 0.5630830966790886,
|
| 194 |
+
"brier": 0.2954194093899411,
|
| 195 |
+
"ece": 0.02856345933167103,
|
| 196 |
+
"aurc": 0.06778106632931055,
|
| 197 |
+
"score_mae": 0.5060575584353655,
|
| 198 |
+
"coverage@0.7": 0.6797436698968428,
|
| 199 |
+
"acc_on_covered@0.7": 0.9059094044607956
|
| 200 |
+
},
|
| 201 |
+
"test_per_domain": {
|
| 202 |
+
"agentic": 0.821117397177404,
|
| 203 |
+
"classification": 0.6115,
|
| 204 |
+
"code": 0.9108344674425007,
|
| 205 |
+
"guardrails": 0.7984073149310548,
|
| 206 |
+
"intents": 0.8505,
|
| 207 |
+
"long_context": 1.0,
|
| 208 |
+
"policy": 0.8061818181818182,
|
| 209 |
+
"reasoning": 0.7180877154615182,
|
| 210 |
+
"support": 0.9815181518151815,
|
| 211 |
+
"tev1_benchmark": 0.6541666666666667,
|
| 212 |
+
"workflows": 0.773
|
| 213 |
+
},
|
| 214 |
+
"test_per_task": [
|
| 215 |
+
{
|
| 216 |
+
"task": "ag_news/topic",
|
| 217 |
+
"n": 500,
|
| 218 |
+
"acc": 0.9,
|
| 219 |
+
"chance": 0.25,
|
| 220 |
+
"nll": 0.2674782864307199,
|
| 221 |
+
"brier": 0.14296993909824685,
|
| 222 |
+
"heldout": false,
|
| 223 |
+
"domain": "classification",
|
| 224 |
+
"ece": 0.05165432274341582,
|
| 225 |
+
"proof_v2 (card)": NaN,
|
| 226 |
+
"\u0394 vs proof_v2": NaN
|
| 227 |
+
},
|
| 228 |
+
{
|
| 229 |
+
"task": "agentharm/refuse",
|
| 230 |
+
"n": 416,
|
| 231 |
+
"acc": 0.6995192307692307,
|
| 232 |
+
"chance": 0.5,
|
| 233 |
+
"nll": 0.6008703433729422,
|
| 234 |
+
"brier": 0.4076762512706606,
|
| 235 |
+
"heldout": true,
|
| 236 |
+
"domain": "guardrails",
|
| 237 |
+
"ece": 0.08916855761064935,
|
| 238 |
+
"proof_v2 (card)": NaN,
|
| 239 |
+
"\u0394 vs proof_v2": NaN
|
| 240 |
+
},
|
| 241 |
+
{
|
| 242 |
+
"task": "agenttrek/finish_now",
|
| 243 |
+
"n": 151,
|
| 244 |
+
"acc": 0.7880794701986755,
|
| 245 |
+
"chance": 0.5,
|
| 246 |
+
"nll": 0.445978463758558,
|
| 247 |
+
"brier": 0.30066953632150406,
|
| 248 |
+
"heldout": false,
|
| 249 |
+
"domain": "agentic",
|
| 250 |
+
"ece": 0.07512128787324918,
|
| 251 |
+
"proof_v2 (card)": NaN,
|
| 252 |
+
"\u0394 vs proof_v2": NaN
|
| 253 |
+
},
|
| 254 |
+
{
|
| 255 |
+
"task": "agenttrek/next_action_type",
|
| 256 |
+
"n": 487,
|
| 257 |
+
"acc": 0.864476386036961,
|
| 258 |
+
"chance": 0.19054952576513148,
|
| 259 |
+
"nll": 0.3929848723152591,
|
| 260 |
+
"brier": 0.2184262054212701,
|
| 261 |
+
"heldout": false,
|
| 262 |
+
"domain": "agentic",
|
| 263 |
+
"ece": 0.09202881025827395,
|
| 264 |
+
"proof_v2 (card)": NaN,
|
| 265 |
+
"\u0394 vs proof_v2": NaN
|
| 266 |
+
},
|
| 267 |
+
{
|
| 268 |
+
"task": "anli/nli",
|
| 269 |
+
"n": 498,
|
| 270 |
+
"acc": 0.4859437751004016,
|
| 271 |
+
"chance": 0.33333333333333326,
|
| 272 |
+
"nll": 1.069221544292677,
|
| 273 |
+
"brier": 0.6544603492386665,
|
| 274 |
+
"heldout": false,
|
| 275 |
+
"domain": "reasoning",
|
| 276 |
+
"ece": 0.14589204983299517,
|
| 277 |
+
"proof_v2 (card)": NaN,
|
| 278 |
+
"\u0394 vs proof_v2": NaN
|
| 279 |
+
},
|
| 280 |
+
{
|
| 281 |
+
"task": "aqua_rat/math",
|
| 282 |
+
"n": 247,
|
| 283 |
+
"acc": 0.340080971659919,
|
| 284 |
+
"chance": 0.2,
|
| 285 |
+
"nll": 1.54525283087603,
|
| 286 |
+
"brier": 0.7739737508966391,
|
| 287 |
+
"heldout": false,
|
| 288 |
+
"domain": "reasoning",
|
| 289 |
+
"ece": 0.080350092668765,
|
| 290 |
+
"proof_v2 (card)": NaN,
|
| 291 |
+
"\u0394 vs proof_v2": NaN
|
| 292 |
+
},
|
| 293 |
+
{
|
| 294 |
+
"task": "arc_challenge/mcq",
|
| 295 |
+
"n": 500,
|
| 296 |
+
"acc": 0.49,
|
| 297 |
+
"chance": 0.25006666666666666,
|
| 298 |
+
"nll": 1.3021090898208785,
|
| 299 |
+
"brier": 0.6752390805252713,
|
| 300 |
+
"heldout": true,
|
| 301 |
+
"domain": "reasoning",
|
| 302 |
+
"ece": 0.1680490539073944,
|
| 303 |
+
"proof_v2 (card)": 0.308,
|
| 304 |
+
"\u0394 vs proof_v2": 0.182
|
| 305 |
+
},
|
| 306 |
+
{
|
| 307 |
+
"task": "arc_easy/mcq",
|
| 308 |
+
"n": 500,
|
| 309 |
+
"acc": 0.612,
|
| 310 |
+
"chance": 0.24980000000000002,
|
| 311 |
+
"nll": 0.9830388274090365,
|
| 312 |
+
"brier": 0.5264415342806388,
|
| 313 |
+
"heldout": true,
|
| 314 |
+
"domain": "reasoning",
|
| 315 |
+
"ece": 0.1215095258355141,
|
| 316 |
+
"proof_v2 (card)": 0.425,
|
| 317 |
+
"\u0394 vs proof_v2": 0.187
|
| 318 |
+
},
|
| 319 |
+
{
|
| 320 |
+
"task": "banking77/intent",
|
| 321 |
+
"n": 500,
|
| 322 |
+
"acc": 0.922,
|
| 323 |
+
"chance": 0.3214666666666666,
|
| 324 |
+
"nll": 0.23128961177584006,
|
| 325 |
+
"brier": 0.11756106243365176,
|
| 326 |
+
"heldout": true,
|
| 327 |
+
"domain": "intents",
|
| 328 |
+
"ece": 0.037492330908775316,
|
| 329 |
+
"proof_v2 (card)": 0.883,
|
| 330 |
+
"\u0394 vs proof_v2": 0.039000000000000035
|
| 331 |
+
},
|
| 332 |
+
{
|
| 333 |
+
"task": "banking77/intent_77",
|
| 334 |
+
"n": 300,
|
| 335 |
+
"acc": 0.58,
|
| 336 |
+
"chance": 0.012987012987012986,
|
| 337 |
+
"nll": 2.159930980304489,
|
| 338 |
+
"brier": 0.6431857202493357,
|
| 339 |
+
"heldout": true,
|
| 340 |
+
"domain": "intents",
|
| 341 |
+
"ece": 0.24079640249411272,
|
| 342 |
+
"proof_v2 (card)": NaN,
|
| 343 |
+
"\u0394 vs proof_v2": NaN
|
| 344 |
+
},
|
| 345 |
+
{
|
| 346 |
+
"task": "bigclonebench/clone",
|
| 347 |
+
"n": 500,
|
| 348 |
+
"acc": 0.962,
|
| 349 |
+
"chance": 0.5,
|
| 350 |
+
"nll": 0.1312969167594565,
|
| 351 |
+
"brier": 0.06444151702850223,
|
| 352 |
+
"heldout": false,
|
| 353 |
+
"domain": "code",
|
| 354 |
+
"ece": 0.03135248208045959,
|
| 355 |
+
"proof_v2 (card)": 0.383,
|
| 356 |
+
"\u0394 vs proof_v2": 0.579
|
| 357 |
+
},
|
| 358 |
+
{
|
| 359 |
+
"task": "bitext/category",
|
| 360 |
+
"n": 500,
|
| 361 |
+
"acc": 1.0,
|
| 362 |
+
"chance": 0.16128124098124094,
|
| 363 |
+
"nll": 0.0013673773490934594,
|
| 364 |
+
"brier": 0.0003828941821233322,
|
| 365 |
+
"heldout": false,
|
| 366 |
+
"domain": "support",
|
| 367 |
+
"ece": 0.001247308015823329,
|
| 368 |
+
"proof_v2 (card)": NaN,
|
| 369 |
+
"\u0394 vs proof_v2": NaN
|
| 370 |
+
},
|
| 371 |
+
{
|
| 372 |
+
"task": "bitext/route",
|
| 373 |
+
"n": 500,
|
| 374 |
+
"acc": 1.0,
|
| 375 |
+
"chance": 0.19016313131313134,
|
| 376 |
+
"nll": 0.00031508885690954,
|
| 377 |
+
"brier": 4.903548307607296e-06,
|
| 378 |
+
"heldout": false,
|
| 379 |
+
"domain": "support",
|
| 380 |
+
"ece": 0.0003137041330337764,
|
| 381 |
+
"proof_v2 (card)": 0.958,
|
| 382 |
+
"\u0394 vs proof_v2": 0.04200000000000004
|
| 383 |
+
},
|
| 384 |
+
{
|
| 385 |
+
"task": "boolq/yes_no",
|
| 386 |
+
"n": 500,
|
| 387 |
+
"acc": 0.822,
|
| 388 |
+
"chance": 0.5,
|
| 389 |
+
"nll": 0.4842173567947466,
|
| 390 |
+
"brier": 0.2840138305161832,
|
| 391 |
+
"heldout": false,
|
| 392 |
+
"domain": "reasoning",
|
| 393 |
+
"ece": 0.0795972980260849,
|
| 394 |
+
"proof_v2 (card)": 0.717,
|
| 395 |
+
"\u0394 vs proof_v2": 0.10499999999999998
|
| 396 |
+
},
|
| 397 |
+
{
|
| 398 |
+
"task": "civil_comments/toxic",
|
| 399 |
+
"n": 500,
|
| 400 |
+
"acc": 0.926,
|
| 401 |
+
"chance": 0.5,
|
| 402 |
+
"nll": 0.2234244500696659,
|
| 403 |
+
"brier": 0.1209681824839501,
|
| 404 |
+
"heldout": false,
|
| 405 |
+
"domain": "guardrails",
|
| 406 |
+
"ece": 0.05123518884181974,
|
| 407 |
+
"proof_v2 (card)": NaN,
|
| 408 |
+
"\u0394 vs proof_v2": NaN
|
| 409 |
+
},
|
| 410 |
+
{
|
| 411 |
+
"task": "clinc150/intent",
|
| 412 |
+
"n": 500,
|
| 413 |
+
"acc": 0.97,
|
| 414 |
+
"chance": 0.12404474984829318,
|
| 415 |
+
"nll": 0.09906445816022005,
|
| 416 |
+
"brier": 0.04996487012600131,
|
| 417 |
+
"heldout": false,
|
| 418 |
+
"domain": "intents",
|
| 419 |
+
"ece": 0.014944501101970714,
|
| 420 |
+
"proof_v2 (card)": 0.85,
|
| 421 |
+
"\u0394 vs proof_v2": 0.12
|
| 422 |
+
},
|
| 423 |
+
{
|
| 424 |
+
"task": "codexglue/code_to_doc",
|
| 425 |
+
"n": 504,
|
| 426 |
+
"acc": 0.9900793650793651,
|
| 427 |
+
"chance": 0.26233465608465617,
|
| 428 |
+
"nll": 0.032474144255576765,
|
| 429 |
+
"brier": 0.016158895201785088,
|
| 430 |
+
"heldout": false,
|
| 431 |
+
"domain": "code",
|
| 432 |
+
"ece": 0.009654374349685009,
|
| 433 |
+
"proof_v2 (card)": 0.969,
|
| 434 |
+
"\u0394 vs proof_v2": 0.021079365079365142
|
| 435 |
+
},
|
| 436 |
+
{
|
| 437 |
+
"task": "codexglue/doc_to_code",
|
| 438 |
+
"n": 504,
|
| 439 |
+
"acc": 0.9761904761904762,
|
| 440 |
+
"chance": 0.2786044973544973,
|
| 441 |
+
"nll": 0.05970113779317181,
|
| 442 |
+
"brier": 0.03072688989396561,
|
| 443 |
+
"heldout": false,
|
| 444 |
+
"domain": "code",
|
| 445 |
+
"ece": 0.012547755761752008,
|
| 446 |
+
"proof_v2 (card)": 0.961,
|
| 447 |
+
"\u0394 vs proof_v2": 0.015190476190476199
|
| 448 |
+
},
|
| 449 |
+
{
|
| 450 |
+
"task": "codexglue/func_name",
|
| 451 |
+
"n": 467,
|
| 452 |
+
"acc": 0.9550321199143469,
|
| 453 |
+
"chance": 0.25499643112062814,
|
| 454 |
+
"nll": 0.14961532151979784,
|
| 455 |
+
"brier": 0.07053876912275295,
|
| 456 |
+
"heldout": false,
|
| 457 |
+
"domain": "code",
|
| 458 |
+
"ece": 0.03948740715388316,
|
| 459 |
+
"proof_v2 (card)": 0.901,
|
| 460 |
+
"\u0394 vs proof_v2": 0.054032119914346866
|
| 461 |
+
},
|
| 462 |
+
{
|
| 463 |
+
"task": "codexglue/lang_id",
|
| 464 |
+
"n": 504,
|
| 465 |
+
"acc": 1.0,
|
| 466 |
+
"chance": 0.2354166666666667,
|
| 467 |
+
"nll": 0.0014352430686462487,
|
| 468 |
+
"brier": 1.2864369183890018e-05,
|
| 469 |
+
"heldout": false,
|
| 470 |
+
"domain": "code",
|
| 471 |
+
"ece": 0.0014309174721203188,
|
| 472 |
+
"proof_v2 (card)": 0.997,
|
| 473 |
+
"\u0394 vs proof_v2": 0.0030000000000000027
|
| 474 |
+
},
|
| 475 |
+
{
|
| 476 |
+
"task": "commonsense_qa/mcq",
|
| 477 |
+
"n": 493,
|
| 478 |
+
"acc": 0.6430020283975659,
|
| 479 |
+
"chance": 0.19999999999999998,
|
| 480 |
+
"nll": 0.9523430099929484,
|
| 481 |
+
"brier": 0.49435087368161623,
|
| 482 |
+
"heldout": false,
|
| 483 |
+
"domain": "reasoning",
|
| 484 |
+
"ece": 0.14308575696442233,
|
| 485 |
+
"proof_v2 (card)": 0.442,
|
| 486 |
+
"\u0394 vs proof_v2": 0.20100202839756592
|
| 487 |
+
},
|
| 488 |
+
{
|
| 489 |
+
"task": "counsel/critique_quality",
|
| 490 |
+
"n": 201,
|
| 491 |
+
"acc": 0.6268656716417911,
|
| 492 |
+
"chance": 0.3333333333333333,
|
| 493 |
+
"nll": 1.5466346353110312,
|
| 494 |
+
"brier": 0.6427543071457474,
|
| 495 |
+
"heldout": false,
|
| 496 |
+
"domain": "agentic",
|
| 497 |
+
"ece": 0.27062572442477023,
|
| 498 |
+
"proof_v2 (card)": NaN,
|
| 499 |
+
"\u0394 vs proof_v2": NaN
|
| 500 |
+
},
|
| 501 |
+
{
|
| 502 |
+
"task": "counsel/step_has_error",
|
| 503 |
+
"n": 201,
|
| 504 |
+
"acc": 0.8308457711442786,
|
| 505 |
+
"chance": 0.5,
|
| 506 |
+
"nll": 0.8236887436017318,
|
| 507 |
+
"brier": 0.3201374357369142,
|
| 508 |
+
"heldout": false,
|
| 509 |
+
"domain": "agentic",
|
| 510 |
+
"ece": 0.14827988011326954,
|
| 511 |
+
"proof_v2 (card)": NaN,
|
| 512 |
+
"\u0394 vs proof_v2": NaN
|
| 513 |
+
},
|
| 514 |
+
{
|
| 515 |
+
"task": "devign/vulnerability",
|
| 516 |
+
"n": 500,
|
| 517 |
+
"acc": 0.632,
|
| 518 |
+
"chance": 0.5,
|
| 519 |
+
"nll": 0.6074751503933221,
|
| 520 |
+
"brier": 0.42706214521965163,
|
| 521 |
+
"heldout": false,
|
| 522 |
+
"domain": "code",
|
| 523 |
+
"ece": 0.07018342161178594,
|
| 524 |
+
"proof_v2 (card)": 0.542,
|
| 525 |
+
"\u0394 vs proof_v2": 0.08999999999999997
|
| 526 |
+
},
|
| 527 |
+
{
|
| 528 |
+
"task": "emotion/6way",
|
| 529 |
+
"n": 500,
|
| 530 |
+
"acc": 0.472,
|
| 531 |
+
"chance": 0.16666666666666663,
|
| 532 |
+
"nll": 1.561316128242761,
|
| 533 |
+
"brier": 0.7356062387487318,
|
| 534 |
+
"heldout": true,
|
| 535 |
+
"domain": "classification",
|
| 536 |
+
"ece": 0.21054179659485817,
|
| 537 |
+
"proof_v2 (card)": NaN,
|
| 538 |
+
"\u0394 vs proof_v2": NaN
|
| 539 |
+
},
|
| 540 |
+
{
|
| 541 |
+
"task": "gsm8k/math",
|
| 542 |
+
"n": 500,
|
| 543 |
+
"acc": 0.7,
|
| 544 |
+
"chance": 0.25,
|
| 545 |
+
"nll": 0.6923719964642078,
|
| 546 |
+
"brier": 0.4014884319624292,
|
| 547 |
+
"heldout": false,
|
| 548 |
+
"domain": "reasoning",
|
| 549 |
+
"ece": 0.08541783547401426,
|
| 550 |
+
"proof_v2 (card)": 0.275,
|
| 551 |
+
"\u0394 vs proof_v2": 0.42499999999999993
|
| 552 |
+
},
|
| 553 |
+
{
|
| 554 |
+
"task": "hellaswag/continuation",
|
| 555 |
+
"n": 500,
|
| 556 |
+
"acc": 0.576,
|
| 557 |
+
"chance": 0.25,
|
| 558 |
+
"nll": 0.968790470642969,
|
| 559 |
+
"brier": 0.5288133586177948,
|
| 560 |
+
"heldout": false,
|
| 561 |
+
"domain": "reasoning",
|
| 562 |
+
"ece": 0.07543198531866073,
|
| 563 |
+
"proof_v2 (card)": NaN,
|
| 564 |
+
"\u0394 vs proof_v2": NaN
|
| 565 |
+
},
|
| 566 |
+
{
|
| 567 |
+
"task": "hotpotqa/comparison_yes_no",
|
| 568 |
+
"n": 26,
|
| 569 |
+
"acc": 0.9230769230769231,
|
| 570 |
+
"chance": 0.5,
|
| 571 |
+
"nll": 0.3400710394176153,
|
| 572 |
+
"brier": 0.15484679762361436,
|
| 573 |
+
"heldout": false,
|
| 574 |
+
"domain": "agentic",
|
| 575 |
+
"ece": 0.08232885369887719,
|
| 576 |
+
"proof_v2 (card)": NaN,
|
| 577 |
+
"\u0394 vs proof_v2": NaN
|
| 578 |
+
},
|
| 579 |
+
{
|
| 580 |
+
"task": "hotpotqa/retrieve",
|
| 581 |
+
"n": 497,
|
| 582 |
+
"acc": 0.8933601609657947,
|
| 583 |
+
"chance": 0.16741720162243304,
|
| 584 |
+
"nll": 0.3457089332123268,
|
| 585 |
+
"brier": 0.16216258689991922,
|
| 586 |
+
"heldout": false,
|
| 587 |
+
"domain": "agentic",
|
| 588 |
+
"ece": 0.04268557528854612,
|
| 589 |
+
"proof_v2 (card)": NaN,
|
| 590 |
+
"\u0394 vs proof_v2": NaN
|
| 591 |
+
},
|
| 592 |
+
{
|
| 593 |
+
"task": "humaneval/completion",
|
| 594 |
+
"n": 119,
|
| 595 |
+
"acc": 0.8487394957983193,
|
| 596 |
+
"chance": 0.39355742296918783,
|
| 597 |
+
"nll": 0.46361927195851294,
|
| 598 |
+
"brier": 0.2730045795813753,
|
| 599 |
+
"heldout": true,
|
| 600 |
+
"domain": "code",
|
| 601 |
+
"ece": 0.13989568708323633,
|
| 602 |
+
"proof_v2 (card)": 0.575,
|
| 603 |
+
"\u0394 vs proof_v2": 0.27373949579831935
|
| 604 |
+
},
|
| 605 |
+
{
|
| 606 |
+
"task": "jailbreak/detect",
|
| 607 |
+
"n": 262,
|
| 608 |
+
"acc": 0.9732824427480916,
|
| 609 |
+
"chance": 0.5,
|
| 610 |
+
"nll": 0.08469655378148465,
|
| 611 |
+
"brier": 0.044003931208438284,
|
| 612 |
+
"heldout": false,
|
| 613 |
+
"domain": "guardrails",
|
| 614 |
+
"ece": 0.030764950368240628,
|
| 615 |
+
"proof_v2 (card)": NaN,
|
| 616 |
+
"\u0394 vs proof_v2": NaN
|
| 617 |
+
},
|
| 618 |
+
{
|
| 619 |
+
"task": "long/contract_clause",
|
| 620 |
+
"n": 150,
|
| 621 |
+
"acc": 1.0,
|
| 622 |
+
"chance": 0.19999999999999996,
|
| 623 |
+
"nll": 0.0002789537957808837,
|
| 624 |
+
"brier": 4.274148862877054e-07,
|
| 625 |
+
"heldout": false,
|
| 626 |
+
"domain": "long_context",
|
| 627 |
+
"ece": 0.00027882695198055973,
|
| 628 |
+
"proof_v2 (card)": NaN,
|
| 629 |
+
"\u0394 vs proof_v2": NaN
|
| 630 |
+
},
|
| 631 |
+
{
|
| 632 |
+
"task": "long/email_thread",
|
| 633 |
+
"n": 150,
|
| 634 |
+
"acc": 1.0,
|
| 635 |
+
"chance": 0.25,
|
| 636 |
+
"nll": 0.0003266528345905802,
|
| 637 |
+
"brier": 2.717841913247229e-06,
|
| 638 |
+
"heldout": false,
|
| 639 |
+
"domain": "long_context",
|
| 640 |
+
"ece": 0.0003259424368540209,
|
| 641 |
+
"proof_v2 (card)": NaN,
|
| 642 |
+
"\u0394 vs proof_v2": NaN
|
| 643 |
+
},
|
| 644 |
+
{
|
| 645 |
+
"task": "long/service_log",
|
| 646 |
+
"n": 150,
|
| 647 |
+
"acc": 1.0,
|
| 648 |
+
"chance": 0.19999999999999996,
|
| 649 |
+
"nll": 0.0010035930563268873,
|
| 650 |
+
"brier": 2.0987077842670875e-06,
|
| 651 |
+
"heldout": false,
|
| 652 |
+
"domain": "long_context",
|
| 653 |
+
"ece": 0.0010028688112894146,
|
| 654 |
+
"proof_v2 (card)": NaN,
|
| 655 |
+
"\u0394 vs proof_v2": NaN
|
| 656 |
+
},
|
| 657 |
+
{
|
| 658 |
+
"task": "massive_en/intent",
|
| 659 |
+
"n": 500,
|
| 660 |
+
"acc": 0.93,
|
| 661 |
+
"chance": 0.12958727789085603,
|
| 662 |
+
"nll": 0.20636169074449573,
|
| 663 |
+
"brier": 0.10375277943487987,
|
| 664 |
+
"heldout": false,
|
| 665 |
+
"domain": "intents",
|
| 666 |
+
"ece": 0.0284088468849659,
|
| 667 |
+
"proof_v2 (card)": NaN,
|
| 668 |
+
"\u0394 vs proof_v2": NaN
|
| 669 |
+
},
|
| 670 |
+
{
|
| 671 |
+
"task": "mbpp/bugspot",
|
| 672 |
+
"n": 256,
|
| 673 |
+
"acc": 0.85546875,
|
| 674 |
+
"chance": 0.40559895833333326,
|
| 675 |
+
"nll": 0.3667814989712497,
|
| 676 |
+
"brier": 0.21843909652329055,
|
| 677 |
+
"heldout": false,
|
| 678 |
+
"domain": "code",
|
| 679 |
+
"ece": 0.037241218378767364,
|
| 680 |
+
"proof_v2 (card)": 0.475,
|
| 681 |
+
"\u0394 vs proof_v2": 0.38046875
|
| 682 |
+
},
|
| 683 |
+
{
|
| 684 |
+
"task": "mbpp/solution",
|
| 685 |
+
"n": 500,
|
| 686 |
+
"acc": 0.978,
|
| 687 |
+
"chance": 0.25,
|
| 688 |
+
"nll": 0.05863042902501184,
|
| 689 |
+
"brier": 0.02981698831494921,
|
| 690 |
+
"heldout": false,
|
| 691 |
+
"domain": "code",
|
| 692 |
+
"ece": 0.01960155099630357,
|
| 693 |
+
"proof_v2 (card)": 0.992,
|
| 694 |
+
"\u0394 vs proof_v2": -0.014000000000000012
|
| 695 |
+
},
|
| 696 |
+
{
|
| 697 |
+
"task": "mmlu/mcq",
|
| 698 |
+
"n": 500,
|
| 699 |
+
"acc": 0.39,
|
| 700 |
+
"chance": 0.25,
|
| 701 |
+
"nll": 1.4543649319559335,
|
| 702 |
+
"brier": 0.7702575193705765,
|
| 703 |
+
"heldout": true,
|
| 704 |
+
"domain": "reasoning",
|
| 705 |
+
"ece": 0.15507307499647138,
|
| 706 |
+
"proof_v2 (card)": NaN,
|
| 707 |
+
"\u0394 vs proof_v2": NaN
|
| 708 |
+
},
|
| 709 |
+
{
|
| 710 |
+
"task": "mnli/claim",
|
| 711 |
+
"n": 500,
|
| 712 |
+
"acc": 0.86,
|
| 713 |
+
"chance": 0.33333333333333326,
|
| 714 |
+
"nll": 0.4006310784481466,
|
| 715 |
+
"brier": 0.2131232579488494,
|
| 716 |
+
"heldout": false,
|
| 717 |
+
"domain": "reasoning",
|
| 718 |
+
"ece": 0.081218329668045,
|
| 719 |
+
"proof_v2 (card)": 0.492,
|
| 720 |
+
"\u0394 vs proof_v2": 0.368
|
| 721 |
+
},
|
| 722 |
+
{
|
| 723 |
+
"task": "openbookqa/mcq",
|
| 724 |
+
"n": 500,
|
| 725 |
+
"acc": 0.572,
|
| 726 |
+
"chance": 0.25,
|
| 727 |
+
"nll": 1.1920220527790952,
|
| 728 |
+
"brier": 0.6155284183368397,
|
| 729 |
+
"heldout": false,
|
| 730 |
+
"domain": "reasoning",
|
| 731 |
+
"ece": 0.21775564682483672,
|
| 732 |
+
"proof_v2 (card)": 0.292,
|
| 733 |
+
"\u0394 vs proof_v2": 0.27999999999999997
|
| 734 |
+
},
|
| 735 |
+
{
|
| 736 |
+
"task": "policy/access_control_transfer",
|
| 737 |
+
"n": 500,
|
| 738 |
+
"acc": 1.0,
|
| 739 |
+
"chance": 0.33333333333333326,
|
| 740 |
+
"nll": 0.002006467206090747,
|
| 741 |
+
"brier": 2.4728843418299015e-05,
|
| 742 |
+
"heldout": false,
|
| 743 |
+
"domain": "policy",
|
| 744 |
+
"ece": 0.001999327063560541,
|
| 745 |
+
"proof_v2 (card)": NaN,
|
| 746 |
+
"\u0394 vs proof_v2": NaN
|
| 747 |
+
},
|
| 748 |
+
{
|
| 749 |
+
"task": "policy/count_threshold_transfer",
|
| 750 |
+
"n": 500,
|
| 751 |
+
"acc": 0.88,
|
| 752 |
+
"chance": 0.10537052392052391,
|
| 753 |
+
"nll": 0.319376940273738,
|
| 754 |
+
"brier": 0.1809248790341789,
|
| 755 |
+
"heldout": false,
|
| 756 |
+
"domain": "policy",
|
| 757 |
+
"ece": 0.04758520478010175,
|
| 758 |
+
"proof_v2 (card)": NaN,
|
| 759 |
+
"\u0394 vs proof_v2": NaN
|
| 760 |
+
},
|
| 761 |
+
{
|
| 762 |
+
"task": "policy/free_shipping_transfer",
|
| 763 |
+
"n": 500,
|
| 764 |
+
"acc": 0.938,
|
| 765 |
+
"chance": 0.5,
|
| 766 |
+
"nll": 0.1287872482436942,
|
| 767 |
+
"brier": 0.08322479293591872,
|
| 768 |
+
"heldout": false,
|
| 769 |
+
"domain": "policy",
|
| 770 |
+
"ece": 0.041308605194091734,
|
| 771 |
+
"proof_v2 (card)": NaN,
|
| 772 |
+
"\u0394 vs proof_v2": NaN
|
| 773 |
+
},
|
| 774 |
+
{
|
| 775 |
+
"task": "policy/invoice_overdue_transfer",
|
| 776 |
+
"n": 500,
|
| 777 |
+
"acc": 0.824,
|
| 778 |
+
"chance": 0.5,
|
| 779 |
+
"nll": 0.3965394942490384,
|
| 780 |
+
"brier": 0.2504247577332596,
|
| 781 |
+
"heldout": false,
|
| 782 |
+
"domain": "policy",
|
| 783 |
+
"ece": 0.01985593855381013,
|
| 784 |
+
"proof_v2 (card)": NaN,
|
| 785 |
+
"\u0394 vs proof_v2": NaN
|
| 786 |
+
},
|
| 787 |
+
{
|
| 788 |
+
"task": "policy/invoice_total_transfer",
|
| 789 |
+
"n": 500,
|
| 790 |
+
"acc": 0.474,
|
| 791 |
+
"chance": 0.5,
|
| 792 |
+
"nll": 0.710681935429573,
|
| 793 |
+
"brier": 0.5164436840846652,
|
| 794 |
+
"heldout": false,
|
| 795 |
+
"domain": "policy",
|
| 796 |
+
"ece": 0.08196303224563599,
|
| 797 |
+
"proof_v2 (card)": NaN,
|
| 798 |
+
"\u0394 vs proof_v2": NaN
|
| 799 |
+
},
|
| 800 |
+
{
|
| 801 |
+
"task": "policy/refund_approval_transfer",
|
| 802 |
+
"n": 500,
|
| 803 |
+
"acc": 0.962,
|
| 804 |
+
"chance": 0.33333333333333326,
|
| 805 |
+
"nll": 0.10006511305032836,
|
| 806 |
+
"brier": 0.057969574647252664,
|
| 807 |
+
"heldout": false,
|
| 808 |
+
"domain": "policy",
|
| 809 |
+
"ece": 0.018834196925163266,
|
| 810 |
+
"proof_v2 (card)": NaN,
|
| 811 |
+
"\u0394 vs proof_v2": NaN
|
| 812 |
+
},
|
| 813 |
+
{
|
| 814 |
+
"task": "policy/return_window_transfer",
|
| 815 |
+
"n": 500,
|
| 816 |
+
"acc": 1.0,
|
| 817 |
+
"chance": 0.33333333333333326,
|
| 818 |
+
"nll": 0.02710788446944207,
|
| 819 |
+
"brier": 0.004185976644178241,
|
| 820 |
+
"heldout": false,
|
| 821 |
+
"domain": "policy",
|
| 822 |
+
"ece": 0.02567652928829193,
|
| 823 |
+
"proof_v2 (card)": NaN,
|
| 824 |
+
"\u0394 vs proof_v2": NaN
|
| 825 |
+
},
|
| 826 |
+
{
|
| 827 |
+
"task": "policy/sla_urgency_transfer",
|
| 828 |
+
"n": 500,
|
| 829 |
+
"acc": 0.606,
|
| 830 |
+
"chance": 0.25,
|
| 831 |
+
"nll": 0.8170870101451874,
|
| 832 |
+
"brier": 0.5123369210022903,
|
| 833 |
+
"heldout": false,
|
| 834 |
+
"domain": "policy",
|
| 835 |
+
"ece": 0.13068743020296097,
|
| 836 |
+
"proof_v2 (card)": NaN,
|
| 837 |
+
"\u0394 vs proof_v2": NaN
|
| 838 |
+
},
|
| 839 |
+
{
|
| 840 |
+
"task": "policy/table_compare_transfer",
|
| 841 |
+
"n": 500,
|
| 842 |
+
"acc": 0.928,
|
| 843 |
+
"chance": 0.5,
|
| 844 |
+
"nll": 0.2105378701629961,
|
| 845 |
+
"brier": 0.11898007306418568,
|
| 846 |
+
"heldout": false,
|
| 847 |
+
"domain": "policy",
|
| 848 |
+
"ece": 0.04929606854915622,
|
| 849 |
+
"proof_v2 (card)": NaN,
|
| 850 |
+
"\u0394 vs proof_v2": NaN
|
| 851 |
+
},
|
| 852 |
+
{
|
| 853 |
+
"task": "policy/table_count_transfer",
|
| 854 |
+
"n": 500,
|
| 855 |
+
"acc": 0.324,
|
| 856 |
+
"chance": 0.12114228549228548,
|
| 857 |
+
"nll": 3.24442003638536,
|
| 858 |
+
"brier": 1.1488026091975905,
|
| 859 |
+
"heldout": false,
|
| 860 |
+
"domain": "policy",
|
| 861 |
+
"ece": 0.5407210965156555,
|
| 862 |
+
"proof_v2 (card)": NaN,
|
| 863 |
+
"\u0394 vs proof_v2": NaN
|
| 864 |
+
},
|
| 865 |
+
{
|
| 866 |
+
"task": "policy/table_extreme_transfer",
|
| 867 |
+
"n": 500,
|
| 868 |
+
"acc": 0.932,
|
| 869 |
+
"chance": 0.12210815295815294,
|
| 870 |
+
"nll": 0.268062693382595,
|
| 871 |
+
"brier": 0.11322196827103552,
|
| 872 |
+
"heldout": false,
|
| 873 |
+
"domain": "policy",
|
| 874 |
+
"ece": 0.05259446609020234,
|
| 875 |
+
"proof_v2 (card)": NaN,
|
| 876 |
+
"\u0394 vs proof_v2": NaN
|
| 877 |
+
},
|
| 878 |
+
{
|
| 879 |
+
"task": "prompt_injections/detect",
|
| 880 |
+
"n": 116,
|
| 881 |
+
"acc": 0.5948275862068966,
|
| 882 |
+
"chance": 0.5,
|
| 883 |
+
"nll": 1.1838965486195179,
|
| 884 |
+
"brier": 0.6826645118148511,
|
| 885 |
+
"heldout": true,
|
| 886 |
+
"domain": "guardrails",
|
| 887 |
+
"ece": 0.33272201953263114,
|
| 888 |
+
"proof_v2 (card)": NaN,
|
| 889 |
+
"\u0394 vs proof_v2": NaN
|
| 890 |
+
},
|
| 891 |
+
{
|
| 892 |
+
"task": "qasc/mcq",
|
| 893 |
+
"n": 500,
|
| 894 |
+
"acc": 0.986,
|
| 895 |
+
"chance": 0.125,
|
| 896 |
+
"nll": 0.028506858559036514,
|
| 897 |
+
"brier": 0.017274368003708154,
|
| 898 |
+
"heldout": false,
|
| 899 |
+
"domain": "reasoning",
|
| 900 |
+
"ece": 0.007774531126022318,
|
| 901 |
+
"proof_v2 (card)": NaN,
|
| 902 |
+
"\u0394 vs proof_v2": NaN
|
| 903 |
+
},
|
| 904 |
+
{
|
| 905 |
+
"task": "sciq/mcq",
|
| 906 |
+
"n": 498,
|
| 907 |
+
"acc": 0.9538152610441767,
|
| 908 |
+
"chance": 0.25,
|
| 909 |
+
"nll": 0.1506263591345834,
|
| 910 |
+
"brier": 0.06985463059035983,
|
| 911 |
+
"heldout": false,
|
| 912 |
+
"domain": "reasoning",
|
| 913 |
+
"ece": 0.020590302994452258,
|
| 914 |
+
"proof_v2 (card)": 0.692,
|
| 915 |
+
"\u0394 vs proof_v2": 0.26181526104417674
|
| 916 |
+
},
|
| 917 |
+
{
|
| 918 |
+
"task": "scitail/support",
|
| 919 |
+
"n": 500,
|
| 920 |
+
"acc": 0.962,
|
| 921 |
+
"chance": 0.5,
|
| 922 |
+
"nll": 0.11612921302905306,
|
| 923 |
+
"brier": 0.05990850379879065,
|
| 924 |
+
"heldout": false,
|
| 925 |
+
"domain": "reasoning",
|
| 926 |
+
"ece": 0.03015338957309728,
|
| 927 |
+
"proof_v2 (card)": NaN,
|
| 928 |
+
"\u0394 vs proof_v2": NaN
|
| 929 |
+
},
|
| 930 |
+
{
|
| 931 |
+
"task": "snli/contradicts",
|
| 932 |
+
"n": 500,
|
| 933 |
+
"acc": 0.99,
|
| 934 |
+
"chance": 0.33333333333333326,
|
| 935 |
+
"nll": 0.042843449617153966,
|
| 936 |
+
"brier": 0.017027529190616314,
|
| 937 |
+
"heldout": false,
|
| 938 |
+
"domain": "reasoning",
|
| 939 |
+
"ece": 0.018511961221695038,
|
| 940 |
+
"proof_v2 (card)": 0.892,
|
| 941 |
+
"\u0394 vs proof_v2": 0.09799999999999998
|
| 942 |
+
},
|
| 943 |
+
{
|
| 944 |
+
"task": "snli/must_be_true",
|
| 945 |
+
"n": 500,
|
| 946 |
+
"acc": 0.986,
|
| 947 |
+
"chance": 0.33333333333333326,
|
| 948 |
+
"nll": 0.06753946821717545,
|
| 949 |
+
"brier": 0.02560326154020915,
|
| 950 |
+
"heldout": false,
|
| 951 |
+
"domain": "reasoning",
|
| 952 |
+
"ece": 0.043202445626258815,
|
| 953 |
+
"proof_v2 (card)": 0.908,
|
| 954 |
+
"\u0394 vs proof_v2": 0.07799999999999996
|
| 955 |
+
},
|
| 956 |
+
{
|
| 957 |
+
"task": "snli/nli",
|
| 958 |
+
"n": 988,
|
| 959 |
+
"acc": 0.8947368421052632,
|
| 960 |
+
"chance": 0.33333333333333326,
|
| 961 |
+
"nll": 0.3554420507991845,
|
| 962 |
+
"brier": 0.17963376213248355,
|
| 963 |
+
"heldout": false,
|
| 964 |
+
"domain": "reasoning",
|
| 965 |
+
"ece": 0.10791712172842222,
|
| 966 |
+
"proof_v2 (card)": NaN,
|
| 967 |
+
"\u0394 vs proof_v2": NaN
|
| 968 |
+
},
|
| 969 |
+
{
|
| 970 |
+
"task": "sst5/score",
|
| 971 |
+
"n": 500,
|
| 972 |
+
"acc": 0.406,
|
| 973 |
+
"chance": 0.2,
|
| 974 |
+
"nll": 1.3002369542717933,
|
| 975 |
+
"brier": 0.6888741553408958,
|
| 976 |
+
"heldout": true,
|
| 977 |
+
"domain": "classification",
|
| 978 |
+
"ece": 0.068851743131876,
|
| 979 |
+
"proof_v2 (card)": NaN,
|
| 980 |
+
"\u0394 vs proof_v2": NaN
|
| 981 |
+
},
|
| 982 |
+
{
|
| 983 |
+
"task": "tev1_test/ag_news",
|
| 984 |
+
"n": 150,
|
| 985 |
+
"acc": 0.92,
|
| 986 |
+
"chance": 0.25,
|
| 987 |
+
"nll": 0.2815313008365532,
|
| 988 |
+
"brier": 0.13456139206846243,
|
| 989 |
+
"heldout": true,
|
| 990 |
+
"domain": "tev1_benchmark",
|
| 991 |
+
"ece": 0.09770219445228576,
|
| 992 |
+
"proof_v2 (card)": NaN,
|
| 993 |
+
"\u0394 vs proof_v2": NaN
|
| 994 |
+
},
|
| 995 |
+
{
|
| 996 |
+
"task": "tev1_test/banking77",
|
| 997 |
+
"n": 200,
|
| 998 |
+
"acc": 0.72,
|
| 999 |
+
"chance": 0.17791666666666664,
|
| 1000 |
+
"nll": 0.777236735031438,
|
| 1001 |
+
"brier": 0.3902663054271286,
|
| 1002 |
+
"heldout": true,
|
| 1003 |
+
"domain": "tev1_benchmark",
|
| 1004 |
+
"ece": 0.13721376925706863,
|
| 1005 |
+
"proof_v2 (card)": NaN,
|
| 1006 |
+
"\u0394 vs proof_v2": NaN
|
| 1007 |
+
},
|
| 1008 |
+
{
|
| 1009 |
+
"task": "tev1_test/boolq",
|
| 1010 |
+
"n": 200,
|
| 1011 |
+
"acc": 0.835,
|
| 1012 |
+
"chance": 0.5,
|
| 1013 |
+
"nll": 0.40902234488283284,
|
| 1014 |
+
"brier": 0.24753055362669119,
|
| 1015 |
+
"heldout": true,
|
| 1016 |
+
"domain": "tev1_benchmark",
|
| 1017 |
+
"ece": 0.04955918818712238,
|
| 1018 |
+
"proof_v2 (card)": NaN,
|
| 1019 |
+
"\u0394 vs proof_v2": NaN
|
| 1020 |
+
},
|
| 1021 |
+
{
|
| 1022 |
+
"task": "tev1_test/mnli",
|
| 1023 |
+
"n": 300,
|
| 1024 |
+
"acc": 0.7233333333333334,
|
| 1025 |
+
"chance": 0.3333333333333333,
|
| 1026 |
+
"nll": 0.6736717029288412,
|
| 1027 |
+
"brier": 0.39782655881245704,
|
| 1028 |
+
"heldout": true,
|
| 1029 |
+
"domain": "tev1_benchmark",
|
| 1030 |
+
"ece": 0.07533363938331603,
|
| 1031 |
+
"proof_v2 (card)": NaN,
|
| 1032 |
+
"\u0394 vs proof_v2": NaN
|
| 1033 |
+
},
|
| 1034 |
+
{
|
| 1035 |
+
"task": "tev1_test/policy",
|
| 1036 |
+
"n": 1200,
|
| 1037 |
+
"acc": 0.5291666666666667,
|
| 1038 |
+
"chance": 0.3333333333333333,
|
| 1039 |
+
"nll": 0.8761954319352905,
|
| 1040 |
+
"brier": 0.5462664224307302,
|
| 1041 |
+
"heldout": true,
|
| 1042 |
+
"domain": "tev1_benchmark",
|
| 1043 |
+
"ece": 0.05868758827447889,
|
| 1044 |
+
"proof_v2 (card)": NaN,
|
| 1045 |
+
"\u0394 vs proof_v2": NaN
|
| 1046 |
+
},
|
| 1047 |
+
{
|
| 1048 |
+
"task": "tev1_test/routing",
|
| 1049 |
+
"n": 600,
|
| 1050 |
+
"acc": 0.4583333333333333,
|
| 1051 |
+
"chance": 0.2,
|
| 1052 |
+
"nll": 1.120521725914441,
|
| 1053 |
+
"brier": 0.5789269511485964,
|
| 1054 |
+
"heldout": true,
|
| 1055 |
+
"domain": "tev1_benchmark",
|
| 1056 |
+
"ece": 0.037736288358767814,
|
| 1057 |
+
"proof_v2 (card)": NaN,
|
| 1058 |
+
"\u0394 vs proof_v2": NaN
|
| 1059 |
+
},
|
| 1060 |
+
{
|
| 1061 |
+
"task": "tev1_test/sst5",
|
| 1062 |
+
"n": 150,
|
| 1063 |
+
"acc": 0.3933333333333333,
|
| 1064 |
+
"chance": 0.19999999999999996,
|
| 1065 |
+
"nll": 1.294863009850184,
|
| 1066 |
+
"brier": 0.6810328679494898,
|
| 1067 |
+
"heldout": true,
|
| 1068 |
+
"domain": "tev1_benchmark",
|
| 1069 |
+
"ece": 0.1081918982664744,
|
| 1070 |
+
"proof_v2 (card)": NaN,
|
| 1071 |
+
"\u0394 vs proof_v2": NaN
|
| 1072 |
+
},
|
| 1073 |
+
{
|
| 1074 |
+
"task": "triage/support_email",
|
| 1075 |
+
"n": 505,
|
| 1076 |
+
"acc": 0.9445544554455445,
|
| 1077 |
+
"chance": 0.32333333333333325,
|
| 1078 |
+
"nll": 0.15705669974757994,
|
| 1079 |
+
"brier": 0.07848840114983771,
|
| 1080 |
+
"heldout": false,
|
| 1081 |
+
"domain": "support",
|
| 1082 |
+
"ece": 0.045869263741049465,
|
| 1083 |
+
"proof_v2 (card)": NaN,
|
| 1084 |
+
"\u0394 vs proof_v2": NaN
|
| 1085 |
+
},
|
| 1086 |
+
{
|
| 1087 |
+
"task": "typed_decisions/agent_trace_observability",
|
| 1088 |
+
"n": 500,
|
| 1089 |
+
"acc": 0.734,
|
| 1090 |
+
"chance": 0.3,
|
| 1091 |
+
"nll": 0.8089503426551818,
|
| 1092 |
+
"brier": 0.45125534421065666,
|
| 1093 |
+
"heldout": false,
|
| 1094 |
+
"domain": "workflows",
|
| 1095 |
+
"ece": 0.22166652178764346,
|
| 1096 |
+
"proof_v2 (card)": NaN,
|
| 1097 |
+
"\u0394 vs proof_v2": NaN
|
| 1098 |
+
},
|
| 1099 |
+
{
|
| 1100 |
+
"task": "typed_decisions/customer_service",
|
| 1101 |
+
"n": 500,
|
| 1102 |
+
"acc": 0.764,
|
| 1103 |
+
"chance": 0.28,
|
| 1104 |
+
"nll": 0.7428050636351109,
|
| 1105 |
+
"brier": 0.40177602517263195,
|
| 1106 |
+
"heldout": false,
|
| 1107 |
+
"domain": "workflows",
|
| 1108 |
+
"ece": 0.20972245779633528,
|
| 1109 |
+
"proof_v2 (card)": NaN,
|
| 1110 |
+
"\u0394 vs proof_v2": NaN
|
| 1111 |
+
},
|
| 1112 |
+
{
|
| 1113 |
+
"task": "typed_decisions/invoice_processing",
|
| 1114 |
+
"n": 500,
|
| 1115 |
+
"acc": 0.826,
|
| 1116 |
+
"chance": 0.35,
|
| 1117 |
+
"nll": 0.5732684296518564,
|
| 1118 |
+
"brier": 0.2997603036136731,
|
| 1119 |
+
"heldout": false,
|
| 1120 |
+
"domain": "workflows",
|
| 1121 |
+
"ece": 0.19463082921504973,
|
| 1122 |
+
"proof_v2 (card)": NaN,
|
| 1123 |
+
"\u0394 vs proof_v2": NaN
|
| 1124 |
+
},
|
| 1125 |
+
{
|
| 1126 |
+
"task": "typed_decisions/security_incidents",
|
| 1127 |
+
"n": 500,
|
| 1128 |
+
"acc": 0.768,
|
| 1129 |
+
"chance": 0.34,
|
| 1130 |
+
"nll": 0.7502468670606613,
|
| 1131 |
+
"brier": 0.4238190137148829,
|
| 1132 |
+
"heldout": false,
|
| 1133 |
+
"domain": "workflows",
|
| 1134 |
+
"ece": 0.23508695220947262,
|
| 1135 |
+
"proof_v2 (card)": NaN,
|
| 1136 |
+
"\u0394 vs proof_v2": NaN
|
| 1137 |
+
},
|
| 1138 |
+
{
|
| 1139 |
+
"task": "winogrande/blank",
|
| 1140 |
+
"n": 500,
|
| 1141 |
+
"acc": 0.662,
|
| 1142 |
+
"chance": 0.5,
|
| 1143 |
+
"nll": 0.6651458515003323,
|
| 1144 |
+
"brier": 0.45323706188140617,
|
| 1145 |
+
"heldout": false,
|
| 1146 |
+
"domain": "reasoning",
|
| 1147 |
+
"ece": 0.1159557296037674,
|
| 1148 |
+
"proof_v2 (card)": NaN,
|
| 1149 |
+
"\u0394 vs proof_v2": NaN
|
| 1150 |
+
},
|
| 1151 |
+
{
|
| 1152 |
+
"task": "yelp/score",
|
| 1153 |
+
"n": 500,
|
| 1154 |
+
"acc": 0.668,
|
| 1155 |
+
"chance": 0.2,
|
| 1156 |
+
"nll": 0.7834262755662202,
|
| 1157 |
+
"brier": 0.4423740240858843,
|
| 1158 |
+
"heldout": false,
|
| 1159 |
+
"domain": "classification",
|
| 1160 |
+
"ece": 0.09068749487400055,
|
| 1161 |
+
"proof_v2 (card)": NaN,
|
| 1162 |
+
"\u0394 vs proof_v2": NaN
|
| 1163 |
+
}
|
| 1164 |
+
],
|
| 1165 |
+
"head_to_head": [
|
| 1166 |
+
{
|
| 1167 |
+
"task": "ag_news/topic",
|
| 1168 |
+
"n": 120,
|
| 1169 |
+
"LightDec": 0.85,
|
| 1170 |
+
"proof_v2": 0.45,
|
| 1171 |
+
"\u0394": 0.39999999999999997
|
| 1172 |
+
},
|
| 1173 |
+
{
|
| 1174 |
+
"task": "agentharm/refuse",
|
| 1175 |
+
"n": 120,
|
| 1176 |
+
"LightDec": 0.6666666666666666,
|
| 1177 |
+
"proof_v2": 0.45,
|
| 1178 |
+
"\u0394": 0.21666666666666662
|
| 1179 |
+
},
|
| 1180 |
+
{
|
| 1181 |
+
"task": "agenttrek/finish_now",
|
| 1182 |
+
"n": 120,
|
| 1183 |
+
"LightDec": 0.7833333333333333,
|
| 1184 |
+
"proof_v2": 0.25833333333333336,
|
| 1185 |
+
"\u0394": 0.5249999999999999
|
| 1186 |
+
},
|
| 1187 |
+
{
|
| 1188 |
+
"task": "agenttrek/next_action_type",
|
| 1189 |
+
"n": 120,
|
| 1190 |
+
"LightDec": 0.875,
|
| 1191 |
+
"proof_v2": 0.075,
|
| 1192 |
+
"\u0394": 0.8
|
| 1193 |
+
},
|
| 1194 |
+
{
|
| 1195 |
+
"task": "anli/nli",
|
| 1196 |
+
"n": 120,
|
| 1197 |
+
"LightDec": 0.5916666666666667,
|
| 1198 |
+
"proof_v2": 0.325,
|
| 1199 |
+
"\u0394": 0.26666666666666666
|
| 1200 |
+
},
|
| 1201 |
+
{
|
| 1202 |
+
"task": "aqua_rat/math",
|
| 1203 |
+
"n": 120,
|
| 1204 |
+
"LightDec": 0.35,
|
| 1205 |
+
"proof_v2": 0.24166666666666667,
|
| 1206 |
+
"\u0394": 0.10833333333333331
|
| 1207 |
+
},
|
| 1208 |
+
{
|
| 1209 |
+
"task": "arc_challenge/mcq",
|
| 1210 |
+
"n": 120,
|
| 1211 |
+
"LightDec": 0.5333333333333333,
|
| 1212 |
+
"proof_v2": 0.30833333333333335,
|
| 1213 |
+
"\u0394": 0.22499999999999998
|
| 1214 |
+
},
|
| 1215 |
+
{
|
| 1216 |
+
"task": "arc_easy/mcq",
|
| 1217 |
+
"n": 120,
|
| 1218 |
+
"LightDec": 0.65,
|
| 1219 |
+
"proof_v2": 0.4583333333333333,
|
| 1220 |
+
"\u0394": 0.1916666666666667
|
| 1221 |
+
},
|
| 1222 |
+
{
|
| 1223 |
+
"task": "banking77/intent",
|
| 1224 |
+
"n": 120,
|
| 1225 |
+
"LightDec": 0.925,
|
| 1226 |
+
"proof_v2": 0.8666666666666667,
|
| 1227 |
+
"\u0394": 0.05833333333333335
|
| 1228 |
+
},
|
| 1229 |
+
{
|
| 1230 |
+
"task": "bigclonebench/clone",
|
| 1231 |
+
"n": 120,
|
| 1232 |
+
"LightDec": 0.9666666666666667,
|
| 1233 |
+
"proof_v2": 0.24166666666666667,
|
| 1234 |
+
"\u0394": 0.725
|
| 1235 |
+
},
|
| 1236 |
+
{
|
| 1237 |
+
"task": "bitext/category",
|
| 1238 |
+
"n": 120,
|
| 1239 |
+
"LightDec": 1.0,
|
| 1240 |
+
"proof_v2": 0.8583333333333333,
|
| 1241 |
+
"\u0394": 0.14166666666666672
|
| 1242 |
+
},
|
| 1243 |
+
{
|
| 1244 |
+
"task": "bitext/route",
|
| 1245 |
+
"n": 120,
|
| 1246 |
+
"LightDec": 1.0,
|
| 1247 |
+
"proof_v2": 0.925,
|
| 1248 |
+
"\u0394": 0.07499999999999996
|
| 1249 |
+
},
|
| 1250 |
+
{
|
| 1251 |
+
"task": "boolq/yes_no",
|
| 1252 |
+
"n": 120,
|
| 1253 |
+
"LightDec": 0.825,
|
| 1254 |
+
"proof_v2": 0.6666666666666666,
|
| 1255 |
+
"\u0394": 0.15833333333333333
|
| 1256 |
+
},
|
| 1257 |
+
{
|
| 1258 |
+
"task": "civil_comments/toxic",
|
| 1259 |
+
"n": 120,
|
| 1260 |
+
"LightDec": 0.9333333333333333,
|
| 1261 |
+
"proof_v2": 0.35833333333333334,
|
| 1262 |
+
"\u0394": 0.575
|
| 1263 |
+
},
|
| 1264 |
+
{
|
| 1265 |
+
"task": "clinc150/intent",
|
| 1266 |
+
"n": 120,
|
| 1267 |
+
"LightDec": 0.975,
|
| 1268 |
+
"proof_v2": 0.6416666666666667,
|
| 1269 |
+
"\u0394": 0.33333333333333326
|
| 1270 |
+
},
|
| 1271 |
+
{
|
| 1272 |
+
"task": "codexglue/code_to_doc",
|
| 1273 |
+
"n": 120,
|
| 1274 |
+
"LightDec": 1.0,
|
| 1275 |
+
"proof_v2": 0.9333333333333333,
|
| 1276 |
+
"\u0394": 0.06666666666666665
|
| 1277 |
+
},
|
| 1278 |
+
{
|
| 1279 |
+
"task": "codexglue/doc_to_code",
|
| 1280 |
+
"n": 120,
|
| 1281 |
+
"LightDec": 0.9833333333333333,
|
| 1282 |
+
"proof_v2": 0.975,
|
| 1283 |
+
"\u0394": 0.008333333333333304
|
| 1284 |
+
},
|
| 1285 |
+
{
|
| 1286 |
+
"task": "codexglue/func_name",
|
| 1287 |
+
"n": 120,
|
| 1288 |
+
"LightDec": 0.9916666666666667,
|
| 1289 |
+
"proof_v2": 0.9166666666666666,
|
| 1290 |
+
"\u0394": 0.07500000000000007
|
| 1291 |
+
},
|
| 1292 |
+
{
|
| 1293 |
+
"task": "codexglue/lang_id",
|
| 1294 |
+
"n": 120,
|
| 1295 |
+
"LightDec": 1.0,
|
| 1296 |
+
"proof_v2": 1.0,
|
| 1297 |
+
"\u0394": 0.0
|
| 1298 |
+
},
|
| 1299 |
+
{
|
| 1300 |
+
"task": "commonsense_qa/mcq",
|
| 1301 |
+
"n": 120,
|
| 1302 |
+
"LightDec": 0.6833333333333333,
|
| 1303 |
+
"proof_v2": 0.4083333333333333,
|
| 1304 |
+
"\u0394": 0.275
|
| 1305 |
+
},
|
| 1306 |
+
{
|
| 1307 |
+
"task": "counsel/critique_quality",
|
| 1308 |
+
"n": 120,
|
| 1309 |
+
"LightDec": 0.6083333333333333,
|
| 1310 |
+
"proof_v2": 0.25833333333333336,
|
| 1311 |
+
"\u0394": 0.3499999999999999
|
| 1312 |
+
},
|
| 1313 |
+
{
|
| 1314 |
+
"task": "counsel/step_has_error",
|
| 1315 |
+
"n": 120,
|
| 1316 |
+
"LightDec": 0.825,
|
| 1317 |
+
"proof_v2": 0.75,
|
| 1318 |
+
"\u0394": 0.07499999999999996
|
| 1319 |
+
},
|
| 1320 |
+
{
|
| 1321 |
+
"task": "devign/vulnerability",
|
| 1322 |
+
"n": 120,
|
| 1323 |
+
"LightDec": 0.6666666666666666,
|
| 1324 |
+
"proof_v2": 0.5666666666666667,
|
| 1325 |
+
"\u0394": 0.09999999999999998
|
| 1326 |
+
},
|
| 1327 |
+
{
|
| 1328 |
+
"task": "emotion/6way",
|
| 1329 |
+
"n": 120,
|
| 1330 |
+
"LightDec": 0.49166666666666664,
|
| 1331 |
+
"proof_v2": 0.36666666666666664,
|
| 1332 |
+
"\u0394": 0.125
|
| 1333 |
+
},
|
| 1334 |
+
{
|
| 1335 |
+
"task": "gsm8k/math",
|
| 1336 |
+
"n": 120,
|
| 1337 |
+
"LightDec": 0.7333333333333333,
|
| 1338 |
+
"proof_v2": 0.125,
|
| 1339 |
+
"\u0394": 0.6083333333333333
|
| 1340 |
+
},
|
| 1341 |
+
{
|
| 1342 |
+
"task": "hellaswag/continuation",
|
| 1343 |
+
"n": 120,
|
| 1344 |
+
"LightDec": 0.6083333333333333,
|
| 1345 |
+
"proof_v2": 0.2916666666666667,
|
| 1346 |
+
"\u0394": 0.3166666666666666
|
| 1347 |
+
},
|
| 1348 |
+
{
|
| 1349 |
+
"task": "hotpotqa/comparison_yes_no",
|
| 1350 |
+
"n": 26,
|
| 1351 |
+
"LightDec": 0.9230769230769231,
|
| 1352 |
+
"proof_v2": 0.46153846153846156,
|
| 1353 |
+
"\u0394": 0.46153846153846156
|
| 1354 |
+
},
|
| 1355 |
+
{
|
| 1356 |
+
"task": "hotpotqa/retrieve",
|
| 1357 |
+
"n": 120,
|
| 1358 |
+
"LightDec": 0.875,
|
| 1359 |
+
"proof_v2": 0.2916666666666667,
|
| 1360 |
+
"\u0394": 0.5833333333333333
|
| 1361 |
+
},
|
| 1362 |
+
{
|
| 1363 |
+
"task": "humaneval/completion",
|
| 1364 |
+
"n": 119,
|
| 1365 |
+
"LightDec": 0.8487394957983193,
|
| 1366 |
+
"proof_v2": 0.5126050420168067,
|
| 1367 |
+
"\u0394": 0.33613445378151263
|
| 1368 |
+
},
|
| 1369 |
+
{
|
| 1370 |
+
"task": "jailbreak/detect",
|
| 1371 |
+
"n": 120,
|
| 1372 |
+
"LightDec": 0.9833333333333333,
|
| 1373 |
+
"proof_v2": 0.5666666666666667,
|
| 1374 |
+
"\u0394": 0.41666666666666663
|
| 1375 |
+
},
|
| 1376 |
+
{
|
| 1377 |
+
"task": "long/contract_clause",
|
| 1378 |
+
"n": 120,
|
| 1379 |
+
"LightDec": 1.0,
|
| 1380 |
+
"proof_v2": 0.23333333333333334,
|
| 1381 |
+
"\u0394": 0.7666666666666666
|
| 1382 |
+
},
|
| 1383 |
+
{
|
| 1384 |
+
"task": "long/email_thread",
|
| 1385 |
+
"n": 120,
|
| 1386 |
+
"LightDec": 1.0,
|
| 1387 |
+
"proof_v2": 0.05,
|
| 1388 |
+
"\u0394": 0.95
|
| 1389 |
+
},
|
| 1390 |
+
{
|
| 1391 |
+
"task": "long/service_log",
|
| 1392 |
+
"n": 120,
|
| 1393 |
+
"LightDec": 1.0,
|
| 1394 |
+
"proof_v2": 0.24166666666666667,
|
| 1395 |
+
"\u0394": 0.7583333333333333
|
| 1396 |
+
},
|
| 1397 |
+
{
|
| 1398 |
+
"task": "massive_en/intent",
|
| 1399 |
+
"n": 120,
|
| 1400 |
+
"LightDec": 0.95,
|
| 1401 |
+
"proof_v2": 0.7416666666666667,
|
| 1402 |
+
"\u0394": 0.20833333333333326
|
| 1403 |
+
},
|
| 1404 |
+
{
|
| 1405 |
+
"task": "mbpp/bugspot",
|
| 1406 |
+
"n": 120,
|
| 1407 |
+
"LightDec": 0.8583333333333333,
|
| 1408 |
+
"proof_v2": 0.575,
|
| 1409 |
+
"\u0394": 0.2833333333333333
|
| 1410 |
+
},
|
| 1411 |
+
{
|
| 1412 |
+
"task": "mbpp/solution",
|
| 1413 |
+
"n": 120,
|
| 1414 |
+
"LightDec": 1.0,
|
| 1415 |
+
"proof_v2": 0.85,
|
| 1416 |
+
"\u0394": 0.15000000000000002
|
| 1417 |
+
},
|
| 1418 |
+
{
|
| 1419 |
+
"task": "mmlu/mcq",
|
| 1420 |
+
"n": 120,
|
| 1421 |
+
"LightDec": 0.4083333333333333,
|
| 1422 |
+
"proof_v2": 0.30833333333333335,
|
| 1423 |
+
"\u0394": 0.09999999999999998
|
| 1424 |
+
},
|
| 1425 |
+
{
|
| 1426 |
+
"task": "mnli/claim",
|
| 1427 |
+
"n": 120,
|
| 1428 |
+
"LightDec": 0.875,
|
| 1429 |
+
"proof_v2": 0.31666666666666665,
|
| 1430 |
+
"\u0394": 0.5583333333333333
|
| 1431 |
+
},
|
| 1432 |
+
{
|
| 1433 |
+
"task": "openbookqa/mcq",
|
| 1434 |
+
"n": 120,
|
| 1435 |
+
"LightDec": 0.65,
|
| 1436 |
+
"proof_v2": 0.2833333333333333,
|
| 1437 |
+
"\u0394": 0.3666666666666667
|
| 1438 |
+
},
|
| 1439 |
+
{
|
| 1440 |
+
"task": "policy/access_control_transfer",
|
| 1441 |
+
"n": 120,
|
| 1442 |
+
"LightDec": 1.0,
|
| 1443 |
+
"proof_v2": 0.23333333333333334,
|
| 1444 |
+
"\u0394": 0.7666666666666666
|
| 1445 |
+
},
|
| 1446 |
+
{
|
| 1447 |
+
"task": "policy/count_threshold_transfer",
|
| 1448 |
+
"n": 120,
|
| 1449 |
+
"LightDec": 0.8666666666666667,
|
| 1450 |
+
"proof_v2": 0.10833333333333334,
|
| 1451 |
+
"\u0394": 0.7583333333333333
|
| 1452 |
+
},
|
| 1453 |
+
{
|
| 1454 |
+
"task": "policy/free_shipping_transfer",
|
| 1455 |
+
"n": 120,
|
| 1456 |
+
"LightDec": 0.925,
|
| 1457 |
+
"proof_v2": 0.7083333333333334,
|
| 1458 |
+
"\u0394": 0.21666666666666667
|
| 1459 |
+
},
|
| 1460 |
+
{
|
| 1461 |
+
"task": "policy/invoice_overdue_transfer",
|
| 1462 |
+
"n": 120,
|
| 1463 |
+
"LightDec": 0.8083333333333333,
|
| 1464 |
+
"proof_v2": 0.6583333333333333,
|
| 1465 |
+
"\u0394": 0.15000000000000002
|
| 1466 |
+
},
|
| 1467 |
+
{
|
| 1468 |
+
"task": "policy/invoice_total_transfer",
|
| 1469 |
+
"n": 120,
|
| 1470 |
+
"LightDec": 0.45,
|
| 1471 |
+
"proof_v2": 0.525,
|
| 1472 |
+
"\u0394": -0.07500000000000001
|
| 1473 |
+
},
|
| 1474 |
+
{
|
| 1475 |
+
"task": "policy/refund_approval_transfer",
|
| 1476 |
+
"n": 120,
|
| 1477 |
+
"LightDec": 0.9583333333333334,
|
| 1478 |
+
"proof_v2": 0.325,
|
| 1479 |
+
"\u0394": 0.6333333333333333
|
| 1480 |
+
},
|
| 1481 |
+
{
|
| 1482 |
+
"task": "policy/return_window_transfer",
|
| 1483 |
+
"n": 120,
|
| 1484 |
+
"LightDec": 1.0,
|
| 1485 |
+
"proof_v2": 0.3416666666666667,
|
| 1486 |
+
"\u0394": 0.6583333333333333
|
| 1487 |
+
},
|
| 1488 |
+
{
|
| 1489 |
+
"task": "policy/sla_urgency_transfer",
|
| 1490 |
+
"n": 120,
|
| 1491 |
+
"LightDec": 0.5166666666666667,
|
| 1492 |
+
"proof_v2": 0.44166666666666665,
|
| 1493 |
+
"\u0394": 0.07500000000000007
|
| 1494 |
+
},
|
| 1495 |
+
{
|
| 1496 |
+
"task": "policy/table_compare_transfer",
|
| 1497 |
+
"n": 120,
|
| 1498 |
+
"LightDec": 0.9166666666666666,
|
| 1499 |
+
"proof_v2": 0.49166666666666664,
|
| 1500 |
+
"\u0394": 0.425
|
| 1501 |
+
},
|
| 1502 |
+
{
|
| 1503 |
+
"task": "policy/table_count_transfer",
|
| 1504 |
+
"n": 120,
|
| 1505 |
+
"LightDec": 0.31666666666666665,
|
| 1506 |
+
"proof_v2": 0.15833333333333333,
|
| 1507 |
+
"\u0394": 0.15833333333333333
|
| 1508 |
+
},
|
| 1509 |
+
{
|
| 1510 |
+
"task": "policy/table_extreme_transfer",
|
| 1511 |
+
"n": 120,
|
| 1512 |
+
"LightDec": 0.925,
|
| 1513 |
+
"proof_v2": 0.14166666666666666,
|
| 1514 |
+
"\u0394": 0.7833333333333334
|
| 1515 |
+
},
|
| 1516 |
+
{
|
| 1517 |
+
"task": "prompt_injections/detect",
|
| 1518 |
+
"n": 116,
|
| 1519 |
+
"LightDec": 0.5948275862068966,
|
| 1520 |
+
"proof_v2": 0.5431034482758621,
|
| 1521 |
+
"\u0394": 0.051724137931034475
|
| 1522 |
+
},
|
| 1523 |
+
{
|
| 1524 |
+
"task": "qasc/mcq",
|
| 1525 |
+
"n": 120,
|
| 1526 |
+
"LightDec": 0.9916666666666667,
|
| 1527 |
+
"proof_v2": 0.8583333333333333,
|
| 1528 |
+
"\u0394": 0.13333333333333341
|
| 1529 |
+
},
|
| 1530 |
+
{
|
| 1531 |
+
"task": "sciq/mcq",
|
| 1532 |
+
"n": 120,
|
| 1533 |
+
"LightDec": 0.9666666666666667,
|
| 1534 |
+
"proof_v2": 0.8666666666666667,
|
| 1535 |
+
"\u0394": 0.09999999999999998
|
| 1536 |
+
},
|
| 1537 |
+
{
|
| 1538 |
+
"task": "scitail/support",
|
| 1539 |
+
"n": 120,
|
| 1540 |
+
"LightDec": 0.9583333333333334,
|
| 1541 |
+
"proof_v2": 0.425,
|
| 1542 |
+
"\u0394": 0.5333333333333334
|
| 1543 |
+
},
|
| 1544 |
+
{
|
| 1545 |
+
"task": "snli/contradicts",
|
| 1546 |
+
"n": 120,
|
| 1547 |
+
"LightDec": 1.0,
|
| 1548 |
+
"proof_v2": 0.35833333333333334,
|
| 1549 |
+
"\u0394": 0.6416666666666666
|
| 1550 |
+
},
|
| 1551 |
+
{
|
| 1552 |
+
"task": "snli/must_be_true",
|
| 1553 |
+
"n": 120,
|
| 1554 |
+
"LightDec": 0.9916666666666667,
|
| 1555 |
+
"proof_v2": 0.925,
|
| 1556 |
+
"\u0394": 0.06666666666666665
|
| 1557 |
+
},
|
| 1558 |
+
{
|
| 1559 |
+
"task": "snli/nli",
|
| 1560 |
+
"n": 120,
|
| 1561 |
+
"LightDec": 0.9083333333333333,
|
| 1562 |
+
"proof_v2": 0.3416666666666667,
|
| 1563 |
+
"\u0394": 0.5666666666666667
|
| 1564 |
+
},
|
| 1565 |
+
{
|
| 1566 |
+
"task": "sst5/score",
|
| 1567 |
+
"n": 120,
|
| 1568 |
+
"LightDec": 0.38333333333333336,
|
| 1569 |
+
"proof_v2": 0.2833333333333333,
|
| 1570 |
+
"\u0394": 0.10000000000000003
|
| 1571 |
+
},
|
| 1572 |
+
{
|
| 1573 |
+
"task": "tev1_test/ag_news",
|
| 1574 |
+
"n": 120,
|
| 1575 |
+
"LightDec": 0.925,
|
| 1576 |
+
"proof_v2": 0.425,
|
| 1577 |
+
"\u0394": 0.5
|
| 1578 |
+
},
|
| 1579 |
+
{
|
| 1580 |
+
"task": "tev1_test/banking77",
|
| 1581 |
+
"n": 120,
|
| 1582 |
+
"LightDec": 0.7166666666666667,
|
| 1583 |
+
"proof_v2": 0.55,
|
| 1584 |
+
"\u0394": 0.16666666666666663
|
| 1585 |
+
},
|
| 1586 |
+
{
|
| 1587 |
+
"task": "tev1_test/boolq",
|
| 1588 |
+
"n": 120,
|
| 1589 |
+
"LightDec": 0.8416666666666667,
|
| 1590 |
+
"proof_v2": 0.5583333333333333,
|
| 1591 |
+
"\u0394": 0.2833333333333333
|
| 1592 |
+
},
|
| 1593 |
+
{
|
| 1594 |
+
"task": "tev1_test/mnli",
|
| 1595 |
+
"n": 120,
|
| 1596 |
+
"LightDec": 0.75,
|
| 1597 |
+
"proof_v2": 0.39166666666666666,
|
| 1598 |
+
"\u0394": 0.35833333333333334
|
| 1599 |
+
},
|
| 1600 |
+
{
|
| 1601 |
+
"task": "tev1_test/policy",
|
| 1602 |
+
"n": 120,
|
| 1603 |
+
"LightDec": 0.5,
|
| 1604 |
+
"proof_v2": 0.4083333333333333,
|
| 1605 |
+
"\u0394": 0.09166666666666667
|
| 1606 |
+
},
|
| 1607 |
+
{
|
| 1608 |
+
"task": "tev1_test/routing",
|
| 1609 |
+
"n": 120,
|
| 1610 |
+
"LightDec": 0.44166666666666665,
|
| 1611 |
+
"proof_v2": 0.2916666666666667,
|
| 1612 |
+
"\u0394": 0.14999999999999997
|
| 1613 |
+
},
|
| 1614 |
+
{
|
| 1615 |
+
"task": "tev1_test/sst5",
|
| 1616 |
+
"n": 120,
|
| 1617 |
+
"LightDec": 0.39166666666666666,
|
| 1618 |
+
"proof_v2": 0.275,
|
| 1619 |
+
"\u0394": 0.11666666666666664
|
| 1620 |
+
},
|
| 1621 |
+
{
|
| 1622 |
+
"task": "triage/support_email",
|
| 1623 |
+
"n": 120,
|
| 1624 |
+
"LightDec": 0.9583333333333334,
|
| 1625 |
+
"proof_v2": 0.39166666666666666,
|
| 1626 |
+
"\u0394": 0.5666666666666667
|
| 1627 |
+
},
|
| 1628 |
+
{
|
| 1629 |
+
"task": "typed_decisions/agent_trace_observability",
|
| 1630 |
+
"n": 120,
|
| 1631 |
+
"LightDec": 0.8083333333333333,
|
| 1632 |
+
"proof_v2": 0.3333333333333333,
|
| 1633 |
+
"\u0394": 0.47500000000000003
|
| 1634 |
+
},
|
| 1635 |
+
{
|
| 1636 |
+
"task": "typed_decisions/customer_service",
|
| 1637 |
+
"n": 120,
|
| 1638 |
+
"LightDec": 0.7416666666666667,
|
| 1639 |
+
"proof_v2": 0.2833333333333333,
|
| 1640 |
+
"\u0394": 0.45833333333333337
|
| 1641 |
+
},
|
| 1642 |
+
{
|
| 1643 |
+
"task": "typed_decisions/invoice_processing",
|
| 1644 |
+
"n": 120,
|
| 1645 |
+
"LightDec": 0.8166666666666667,
|
| 1646 |
+
"proof_v2": 0.475,
|
| 1647 |
+
"\u0394": 0.3416666666666667
|
| 1648 |
+
},
|
| 1649 |
+
{
|
| 1650 |
+
"task": "typed_decisions/security_incidents",
|
| 1651 |
+
"n": 120,
|
| 1652 |
+
"LightDec": 0.7583333333333333,
|
| 1653 |
+
"proof_v2": 0.5,
|
| 1654 |
+
"\u0394": 0.2583333333333333
|
| 1655 |
+
},
|
| 1656 |
+
{
|
| 1657 |
+
"task": "winogrande/blank",
|
| 1658 |
+
"n": 120,
|
| 1659 |
+
"LightDec": 0.6916666666666667,
|
| 1660 |
+
"proof_v2": 0.5583333333333333,
|
| 1661 |
+
"\u0394": 0.1333333333333333
|
| 1662 |
+
},
|
| 1663 |
+
{
|
| 1664 |
+
"task": "yelp/score",
|
| 1665 |
+
"n": 120,
|
| 1666 |
+
"LightDec": 0.625,
|
| 1667 |
+
"proof_v2": 0.325,
|
| 1668 |
+
"\u0394": 0.3
|
| 1669 |
+
}
|
| 1670 |
+
],
|
| 1671 |
+
"head_to_head_error": null,
|
| 1672 |
+
"v3_benchmarks": {
|
| 1673 |
+
"tev1": {
|
| 1674 |
+
"per_task": [
|
| 1675 |
+
{
|
| 1676 |
+
"task": "tev1_test/ag_news",
|
| 1677 |
+
"n": 150,
|
| 1678 |
+
"acc": 0.92,
|
| 1679 |
+
"overlaps_LightDec_training_source": true
|
| 1680 |
+
},
|
| 1681 |
+
{
|
| 1682 |
+
"task": "tev1_test/banking77",
|
| 1683 |
+
"n": 200,
|
| 1684 |
+
"acc": 0.72,
|
| 1685 |
+
"overlaps_LightDec_training_source": false
|
| 1686 |
+
},
|
| 1687 |
+
{
|
| 1688 |
+
"task": "tev1_test/boolq",
|
| 1689 |
+
"n": 200,
|
| 1690 |
+
"acc": 0.835,
|
| 1691 |
+
"overlaps_LightDec_training_source": true
|
| 1692 |
+
},
|
| 1693 |
+
{
|
| 1694 |
+
"task": "tev1_test/mnli",
|
| 1695 |
+
"n": 300,
|
| 1696 |
+
"acc": 0.7233333333333334,
|
| 1697 |
+
"overlaps_LightDec_training_source": true
|
| 1698 |
+
},
|
| 1699 |
+
{
|
| 1700 |
+
"task": "tev1_test/policy",
|
| 1701 |
+
"n": 1200,
|
| 1702 |
+
"acc": 0.5291666666666667,
|
| 1703 |
+
"overlaps_LightDec_training_source": false
|
| 1704 |
+
},
|
| 1705 |
+
{
|
| 1706 |
+
"task": "tev1_test/routing",
|
| 1707 |
+
"n": 600,
|
| 1708 |
+
"acc": 0.4583333333333333,
|
| 1709 |
+
"overlaps_LightDec_training_source": false
|
| 1710 |
+
},
|
| 1711 |
+
{
|
| 1712 |
+
"task": "tev1_test/sst5",
|
| 1713 |
+
"n": 150,
|
| 1714 |
+
"acc": 0.3933333333333333,
|
| 1715 |
+
"overlaps_LightDec_training_source": false
|
| 1716 |
+
}
|
| 1717 |
+
],
|
| 1718 |
+
"all": 0.5839285714285715,
|
| 1719 |
+
"no_overlap": 0.5176744186046511,
|
| 1720 |
+
"n": 2800,
|
| 1721 |
+
"published_tev1_4b": {
|
| 1722 |
+
"main decisions (1,000)": 0.88,
|
| 1723 |
+
"policy transfer (300)": 1.0
|
| 1724 |
+
}
|
| 1725 |
+
},
|
| 1726 |
+
"triage": {
|
| 1727 |
+
"emails": 101,
|
| 1728 |
+
"per_question": {
|
| 1729 |
+
"topic": 0.9900990099009901,
|
| 1730 |
+
"refund": 1.0,
|
| 1731 |
+
"breakage": 0.9900990099009901,
|
| 1732 |
+
"anger": 0.9207920792079208,
|
| 1733 |
+
"judgment": 0.8217821782178217
|
| 1734 |
+
},
|
| 1735 |
+
"pile_accuracy": 0.9405940594059405
|
| 1736 |
+
},
|
| 1737 |
+
"long_context": {
|
| 1738 |
+
"max_len": 2048,
|
| 1739 |
+
"per_task": [
|
| 1740 |
+
{
|
| 1741 |
+
"task": "long/contract_clause",
|
| 1742 |
+
"n": 150,
|
| 1743 |
+
"acc": 1.0,
|
| 1744 |
+
"chance": 0.19999999999999996
|
| 1745 |
+
},
|
| 1746 |
+
{
|
| 1747 |
+
"task": "long/email_thread",
|
| 1748 |
+
"n": 150,
|
| 1749 |
+
"acc": 1.0,
|
| 1750 |
+
"chance": 0.25
|
| 1751 |
+
},
|
| 1752 |
+
{
|
| 1753 |
+
"task": "long/service_log",
|
| 1754 |
+
"n": 150,
|
| 1755 |
+
"acc": 1.0,
|
| 1756 |
+
"chance": 0.19999999999999996
|
| 1757 |
+
}
|
| 1758 |
+
]
|
| 1759 |
+
}
|
| 1760 |
+
},
|
| 1761 |
+
"architecture": "FalconDec",
|
| 1762 |
+
"latency": {
|
| 1763 |
+
"gpu": "NVIDIA RTX PRO 6000 Blackwell Server Edition",
|
| 1764 |
+
"calls": [
|
| 1765 |
+
{
|
| 1766 |
+
"device": "cuda",
|
| 1767 |
+
"weights": "fp16",
|
| 1768 |
+
"questions": 1,
|
| 1769 |
+
"p50_ms": 10.05,
|
| 1770 |
+
"p95_ms": 10.43
|
| 1771 |
+
},
|
| 1772 |
+
{
|
| 1773 |
+
"device": "cuda",
|
| 1774 |
+
"weights": "fp16",
|
| 1775 |
+
"questions": 5,
|
| 1776 |
+
"p50_ms": 11.32,
|
| 1777 |
+
"p95_ms": 11.41
|
| 1778 |
+
},
|
| 1779 |
+
{
|
| 1780 |
+
"device": "cuda",
|
| 1781 |
+
"weights": "fp16",
|
| 1782 |
+
"questions": 10,
|
| 1783 |
+
"p50_ms": 12.57,
|
| 1784 |
+
"p95_ms": 12.69
|
| 1785 |
+
},
|
| 1786 |
+
{
|
| 1787 |
+
"device": "cpu fp32",
|
| 1788 |
+
"weights": "int8 file",
|
| 1789 |
+
"questions": 1,
|
| 1790 |
+
"p50_ms": 48.4,
|
| 1791 |
+
"p95_ms": 48.8
|
| 1792 |
+
}
|
| 1793 |
+
],
|
| 1794 |
+
"throughput_per_s": 2588.8502560238694
|
| 1795 |
+
},
|
| 1796 |
+
"env": {
|
| 1797 |
+
"torch": "2.9.0+cu130",
|
| 1798 |
+
"transformers": "5.17.0",
|
| 1799 |
+
"python": "3.12.12",
|
| 1800 |
+
"gpu": "NVIDIA RTX PRO 6000 Blackwell Server Edition"
|
| 1801 |
+
}
|
| 1802 |
+
}
|
compact-int8/model_int8.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:a93d550e1c95773af46f11eefcde831edd723e694b20d52e127c1b790ff38959
|
| 3 |
+
size 160530522
|
compact-int8/tokenizer/tokenizer.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
compact-int8/tokenizer/tokenizer_config.json
ADDED
|
@@ -0,0 +1,17 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"backend": "tokenizers",
|
| 3 |
+
"clean_up_tokenization_spaces": true,
|
| 4 |
+
"cls_token": "[CLS]",
|
| 5 |
+
"is_local": false,
|
| 6 |
+
"local_files_only": false,
|
| 7 |
+
"mask_token": "[MASK]",
|
| 8 |
+
"model_input_names": [
|
| 9 |
+
"input_ids",
|
| 10 |
+
"attention_mask"
|
| 11 |
+
],
|
| 12 |
+
"model_max_length": 8192,
|
| 13 |
+
"pad_token": "[PAD]",
|
| 14 |
+
"sep_token": "[SEP]",
|
| 15 |
+
"tokenizer_class": "TokenizersBackend",
|
| 16 |
+
"unk_token": "[UNK]"
|
| 17 |
+
}
|
encoder/config.json
ADDED
|
@@ -0,0 +1,79 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"architectures": [
|
| 3 |
+
"ModernBertForMaskedLM"
|
| 4 |
+
],
|
| 5 |
+
"attention_bias": false,
|
| 6 |
+
"attention_dropout": 0.0,
|
| 7 |
+
"bos_token_id": 50281,
|
| 8 |
+
"causal_mask": false,
|
| 9 |
+
"classifier_activation": "gelu",
|
| 10 |
+
"classifier_bias": false,
|
| 11 |
+
"classifier_dropout": 0.0,
|
| 12 |
+
"classifier_pooling": "mean",
|
| 13 |
+
"cls_token_id": 50281,
|
| 14 |
+
"decoder_bias": true,
|
| 15 |
+
"deterministic_flash_attn": false,
|
| 16 |
+
"dtype": "float32",
|
| 17 |
+
"embedding_dropout": 0.0,
|
| 18 |
+
"eos_token_id": 50282,
|
| 19 |
+
"global_attn_every_n_layers": 3,
|
| 20 |
+
"gradient_checkpointing": false,
|
| 21 |
+
"hidden_activation": "gelu",
|
| 22 |
+
"hidden_size": 768,
|
| 23 |
+
"initializer_cutoff_factor": 2.0,
|
| 24 |
+
"initializer_range": 0.02,
|
| 25 |
+
"intermediate_size": 1152,
|
| 26 |
+
"is_causal": false,
|
| 27 |
+
"layer_norm_eps": 1e-05,
|
| 28 |
+
"layer_types": [
|
| 29 |
+
"full_attention",
|
| 30 |
+
"sliding_attention",
|
| 31 |
+
"sliding_attention",
|
| 32 |
+
"full_attention",
|
| 33 |
+
"sliding_attention",
|
| 34 |
+
"sliding_attention",
|
| 35 |
+
"full_attention",
|
| 36 |
+
"sliding_attention",
|
| 37 |
+
"sliding_attention",
|
| 38 |
+
"full_attention",
|
| 39 |
+
"sliding_attention",
|
| 40 |
+
"sliding_attention",
|
| 41 |
+
"full_attention",
|
| 42 |
+
"sliding_attention",
|
| 43 |
+
"sliding_attention",
|
| 44 |
+
"full_attention",
|
| 45 |
+
"sliding_attention",
|
| 46 |
+
"sliding_attention",
|
| 47 |
+
"full_attention",
|
| 48 |
+
"sliding_attention",
|
| 49 |
+
"sliding_attention",
|
| 50 |
+
"full_attention"
|
| 51 |
+
],
|
| 52 |
+
"local_attention": 128,
|
| 53 |
+
"max_position_embeddings": 7999,
|
| 54 |
+
"mlp_bias": false,
|
| 55 |
+
"mlp_dropout": 0.0,
|
| 56 |
+
"model_type": "modernbert",
|
| 57 |
+
"norm_bias": false,
|
| 58 |
+
"norm_eps": 1e-05,
|
| 59 |
+
"num_attention_heads": 12,
|
| 60 |
+
"num_hidden_layers": 22,
|
| 61 |
+
"pad_token_id": 50283,
|
| 62 |
+
"position_embedding_type": "sans_pos",
|
| 63 |
+
"rope_parameters": {
|
| 64 |
+
"full_attention": {
|
| 65 |
+
"rope_theta": 160000.0,
|
| 66 |
+
"rope_type": "default"
|
| 67 |
+
},
|
| 68 |
+
"sliding_attention": {
|
| 69 |
+
"rope_theta": 160000.0,
|
| 70 |
+
"rope_type": "default"
|
| 71 |
+
}
|
| 72 |
+
},
|
| 73 |
+
"sep_token_id": 50282,
|
| 74 |
+
"sparse_pred_ignore_index": -100,
|
| 75 |
+
"sparse_prediction": false,
|
| 76 |
+
"tie_word_embeddings": true,
|
| 77 |
+
"transformers_version": "5.17.0",
|
| 78 |
+
"vocab_size": 50368
|
| 79 |
+
}
|
falcondec_config.json
ADDED
|
@@ -0,0 +1,54 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"max_len": 2048,
|
| 3 |
+
"long_max_len": 2048,
|
| 4 |
+
"long_opts_threshold": 24,
|
| 5 |
+
"head_max_len": 192,
|
| 6 |
+
"max_tok_per_opt": 24,
|
| 7 |
+
"max_opts_single_pass": 96,
|
| 8 |
+
"interact_layers": 2,
|
| 9 |
+
"interact_heads": 8,
|
| 10 |
+
"name": "LightDec_V2_Long",
|
| 11 |
+
"backbone": "jhu-clsp/ettin-encoder-150m",
|
| 12 |
+
"backbone_init": "pretrained",
|
| 13 |
+
"lineage": [
|
| 14 |
+
"jhu-clsp/ettin-encoder-150m"
|
| 15 |
+
],
|
| 16 |
+
"parent_version": null,
|
| 17 |
+
"special": {
|
| 18 |
+
"cls": 50281,
|
| 19 |
+
"sep": 50282,
|
| 20 |
+
"mask": 50284,
|
| 21 |
+
"pad": 50283
|
| 22 |
+
},
|
| 23 |
+
"defer_threshold": 0.7,
|
| 24 |
+
"qtypes": {
|
| 25 |
+
"choice": 0,
|
| 26 |
+
"noul": 1,
|
| 27 |
+
"score": 2
|
| 28 |
+
},
|
| 29 |
+
"n_buckets": 4,
|
| 30 |
+
"version": "1.0.0",
|
| 31 |
+
"notebook_version": "V3.7",
|
| 32 |
+
"created": "2026-09-28T10:52:36.426087Z",
|
| 33 |
+
"temperature": [
|
| 34 |
+
[
|
| 35 |
+
2.37591814994812,
|
| 36 |
+
2.0426506996154785,
|
| 37 |
+
1.6455482244491577,
|
| 38 |
+
1.8417285680770874
|
| 39 |
+
],
|
| 40 |
+
[
|
| 41 |
+
2.4885058403015137,
|
| 42 |
+
2.4885058403015137,
|
| 43 |
+
2.4885058403015137,
|
| 44 |
+
2.4885058403015137
|
| 45 |
+
],
|
| 46 |
+
[
|
| 47 |
+
2.2925639152526855,
|
| 48 |
+
2.2925639152526855,
|
| 49 |
+
2.2925639152526855,
|
| 50 |
+
2.2925639152526855
|
| 51 |
+
]
|
| 52 |
+
],
|
| 53 |
+
"weights": "model.safetensors"
|
| 54 |
+
}
|
falcondec_modeling.py
ADDED
|
@@ -0,0 +1,358 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# -*- coding: utf-8 -*-
|
| 2 |
+
"""FalconDec — Falcon Decision model: single-pass, typed, calibrated closed-set decisions.
|
| 3 |
+
|
| 4 |
+
Layout : [CLS] question [SEP] [MASK] opt_1 ... [MASK] opt_k [SEP] state [SEP]
|
| 5 |
+
Head : marker vectors + CLS context + question-type embedding
|
| 6 |
+
-> permutation-equivariant set transformer (options attend to each other; no positions)
|
| 7 |
+
-> MLP -> one logit per option -> softmax over this question's options
|
| 8 |
+
Calib. : temperature per (question type, option-count bucket), stored in the checkpoint
|
| 9 |
+
"""
|
| 10 |
+
from __future__ import annotations
|
| 11 |
+
|
| 12 |
+
import contextlib
|
| 13 |
+
import json
|
| 14 |
+
import shutil
|
| 15 |
+
from pathlib import Path
|
| 16 |
+
|
| 17 |
+
import numpy as np
|
| 18 |
+
import torch
|
| 19 |
+
import torch.nn as nn
|
| 20 |
+
|
| 21 |
+
QTYPES = {"choice": 0, "noul": 1, "score": 2}
|
| 22 |
+
N_BUCKETS = 4
|
| 23 |
+
CONFIG_FILE = "falcondec_config.json"
|
| 24 |
+
FP16_FILE = "model.safetensors"
|
| 25 |
+
INT8_FILE = "model_int8.safetensors"
|
| 26 |
+
MODEL_KEYS = ("input_ids", "attention_mask", "marker_pos", "marker_mask", "qtype")
|
| 27 |
+
|
| 28 |
+
|
| 29 |
+
def n_bucket(n: int) -> int:
|
| 30 |
+
return 0 if n <= 2 else 1 if n <= 5 else 2 if n <= 12 else 3
|
| 31 |
+
|
| 32 |
+
|
| 33 |
+
def special_ids(tok) -> dict:
|
| 34 |
+
sp = {"cls": tok.cls_token_id, "sep": tok.sep_token_id, "mask": tok.mask_token_id, "pad": tok.pad_token_id}
|
| 35 |
+
if sp["cls"] is None:
|
| 36 |
+
sp["cls"] = tok.bos_token_id
|
| 37 |
+
if sp["sep"] is None:
|
| 38 |
+
sp["sep"] = tok.eos_token_id
|
| 39 |
+
if sp["pad"] is None:
|
| 40 |
+
sp["pad"] = sp["sep"]
|
| 41 |
+
if sp["mask"] is None:
|
| 42 |
+
raise ValueError("FalconDec needs a tokenizer with a mask token (used as the option marker).")
|
| 43 |
+
return {k: int(v) for k, v in sp.items()}
|
| 44 |
+
|
| 45 |
+
|
| 46 |
+
def _as_list(x):
|
| 47 |
+
return x.tolist() if hasattr(x, "tolist") else list(x)
|
| 48 |
+
|
| 49 |
+
|
| 50 |
+
def assemble(q_ids, opt_ids, s_ids, sp, max_len=512, head_max_len=192, max_tok_per_opt=24,
|
| 51 |
+
long_max_len=2048, long_opts_threshold=24):
|
| 52 |
+
"""Build one input sequence. Returns (ids, marker_positions)."""
|
| 53 |
+
n = len(opt_ids)
|
| 54 |
+
eff = max_len if n <= long_opts_threshold else max(max_len, long_max_len)
|
| 55 |
+
q = _as_list(q_ids)[:96]
|
| 56 |
+
need = len(q) + 2 + n * (max_tok_per_opt + 1)
|
| 57 |
+
head = min(eff - 64, max(head_max_len, need))
|
| 58 |
+
per = max(2, min(max_tok_per_opt, (head - len(q) - 2) // max(n, 1) - 1))
|
| 59 |
+
ids = [sp["cls"]] + q + [sp["sep"]]
|
| 60 |
+
markers = []
|
| 61 |
+
for o in opt_ids:
|
| 62 |
+
markers.append(len(ids))
|
| 63 |
+
ids.append(sp["mask"])
|
| 64 |
+
ids.extend(_as_list(o[:per]))
|
| 65 |
+
ids.append(sp["sep"])
|
| 66 |
+
room = eff - len(ids) - 1
|
| 67 |
+
if room > 0 and s_ids is not None and len(s_ids) > 0:
|
| 68 |
+
ids.extend(_as_list(s_ids[:room]))
|
| 69 |
+
ids.append(sp["sep"])
|
| 70 |
+
if len(ids) > eff:
|
| 71 |
+
ids = ids[:eff]
|
| 72 |
+
if markers and markers[-1] >= len(ids):
|
| 73 |
+
raise ValueError("Too many options for one pass; reduce options or raise long_max_len.")
|
| 74 |
+
return ids, markers
|
| 75 |
+
|
| 76 |
+
|
| 77 |
+
def collate_features(feats, pad_id, device=None):
|
| 78 |
+
"""feats: list of (ids, markers, qtype_index) -> dict of padded tensors matching FalconDec.forward."""
|
| 79 |
+
B = len(feats)
|
| 80 |
+
T = max(len(f[0]) for f in feats)
|
| 81 |
+
K = max(len(f[1]) for f in feats)
|
| 82 |
+
ids = torch.full((B, T), pad_id, dtype=torch.long)
|
| 83 |
+
att = torch.zeros((B, T), dtype=torch.long)
|
| 84 |
+
mpos = torch.zeros((B, K), dtype=torch.long)
|
| 85 |
+
mmask = torch.zeros((B, K), dtype=torch.bool)
|
| 86 |
+
qt = torch.zeros(B, dtype=torch.long)
|
| 87 |
+
for i, (x, mk, q) in enumerate(feats):
|
| 88 |
+
ids[i, : len(x)] = torch.as_tensor(x, dtype=torch.long)
|
| 89 |
+
att[i, : len(x)] = 1
|
| 90 |
+
mpos[i, : len(mk)] = torch.as_tensor(mk, dtype=torch.long)
|
| 91 |
+
mmask[i, : len(mk)] = True
|
| 92 |
+
qt[i] = int(q)
|
| 93 |
+
out = dict(input_ids=ids, attention_mask=att, marker_pos=mpos, marker_mask=mmask, qtype=qt)
|
| 94 |
+
if device is not None:
|
| 95 |
+
out = {k: v.to(device, non_blocking=True) for k, v in out.items()}
|
| 96 |
+
return out
|
| 97 |
+
|
| 98 |
+
|
| 99 |
+
class OptionInteraction(nn.Module):
|
| 100 |
+
"""Set transformer over the options of one question (no positional encoding => order-equivariant)."""
|
| 101 |
+
|
| 102 |
+
def __init__(self, d, n_layers=2, n_heads=8, dropout=0.1):
|
| 103 |
+
super().__init__()
|
| 104 |
+
layer = nn.TransformerEncoderLayer(d, n_heads, dim_feedforward=2 * d, dropout=dropout,
|
| 105 |
+
activation="gelu", batch_first=True, norm_first=True)
|
| 106 |
+
self.enc = nn.TransformerEncoder(layer, n_layers, enable_nested_tensor=False)
|
| 107 |
+
|
| 108 |
+
def forward(self, x, mask):
|
| 109 |
+
return self.enc(x, src_key_padding_mask=~mask)
|
| 110 |
+
|
| 111 |
+
|
| 112 |
+
class FalconDec(nn.Module):
|
| 113 |
+
def __init__(self, encoder, fcfg: dict):
|
| 114 |
+
super().__init__()
|
| 115 |
+
self.encoder = encoder
|
| 116 |
+
self.fcfg = dict(fcfg)
|
| 117 |
+
d = encoder.config.hidden_size
|
| 118 |
+
self.qtype_emb = nn.Embedding(len(QTYPES), d)
|
| 119 |
+
self.ctx_proj = nn.Linear(d, d)
|
| 120 |
+
self.opt_norm = nn.LayerNorm(d)
|
| 121 |
+
self.interact = OptionInteraction(d, int(self.fcfg.get("interact_layers", 2)),
|
| 122 |
+
int(self.fcfg.get("interact_heads", 8)))
|
| 123 |
+
self.scorer = nn.Sequential(nn.Linear(d, d), nn.GELU(), nn.Dropout(0.1), nn.Linear(d, 1))
|
| 124 |
+
self.register_buffer("temperature", torch.ones(len(QTYPES), N_BUCKETS))
|
| 125 |
+
|
| 126 |
+
@property
|
| 127 |
+
def device(self):
|
| 128 |
+
return next(self.parameters()).device
|
| 129 |
+
|
| 130 |
+
def num_parameters(self):
|
| 131 |
+
return sum(p.numel() for p in self.parameters())
|
| 132 |
+
|
| 133 |
+
def forward(self, input_ids, attention_mask, marker_pos, marker_mask, qtype):
|
| 134 |
+
h = self.encoder(input_ids=input_ids, attention_mask=attention_mask).last_hidden_state
|
| 135 |
+
B, K = marker_pos.shape
|
| 136 |
+
idx = marker_pos.unsqueeze(-1).expand(B, K, h.size(-1))
|
| 137 |
+
opt = h.gather(1, idx)
|
| 138 |
+
ctx = self.ctx_proj(h[:, 0]).unsqueeze(1)
|
| 139 |
+
x = self.opt_norm(opt + ctx + self.qtype_emb(qtype).unsqueeze(1))
|
| 140 |
+
x = self.interact(x, marker_mask)
|
| 141 |
+
logits = self.scorer(x).squeeze(-1).float()
|
| 142 |
+
return logits.masked_fill(~marker_mask, -1e4)
|
| 143 |
+
|
| 144 |
+
|
| 145 |
+
# ------------------------------------------------------------------ storage
|
| 146 |
+
def clean_state_dict(sd):
|
| 147 |
+
return {k.replace("_orig_mod.", ""): v.detach().cpu().contiguous() for k, v in sd.items()}
|
| 148 |
+
|
| 149 |
+
|
| 150 |
+
def quantize_int8(sd, min_numel=4096):
|
| 151 |
+
"""Per-output-channel symmetric int8 for every matrix; fp16 for everything else."""
|
| 152 |
+
out = {}
|
| 153 |
+
for k, v in sd.items():
|
| 154 |
+
if v.is_floating_point() and v.ndim == 2 and v.numel() >= min_numel:
|
| 155 |
+
w = v.float()
|
| 156 |
+
s = (w.abs().amax(dim=1, keepdim=True) / 127.0).clamp_min(1e-12)
|
| 157 |
+
out[k] = torch.round(w / s).clamp_(-127, 127).to(torch.int8).contiguous()
|
| 158 |
+
out[k + "::scale"] = s.contiguous()
|
| 159 |
+
elif v.is_floating_point():
|
| 160 |
+
out[k] = v.half().contiguous()
|
| 161 |
+
else:
|
| 162 |
+
out[k] = v.contiguous()
|
| 163 |
+
return out
|
| 164 |
+
|
| 165 |
+
|
| 166 |
+
def dequantize_int8(sd):
|
| 167 |
+
out = {}
|
| 168 |
+
for k, v in sd.items():
|
| 169 |
+
if k.endswith("::scale"):
|
| 170 |
+
continue
|
| 171 |
+
s = sd.get(k + "::scale")
|
| 172 |
+
out[k] = (v.float() * s).half() if s is not None else v
|
| 173 |
+
return out
|
| 174 |
+
|
| 175 |
+
|
| 176 |
+
def save_falcondec(model, tok, out_dir, int8=False, extra_files=None):
|
| 177 |
+
from safetensors.torch import save_file
|
| 178 |
+
out = Path(out_dir)
|
| 179 |
+
out.mkdir(parents=True, exist_ok=True)
|
| 180 |
+
enc_cfg = model.encoder.config
|
| 181 |
+
if hasattr(enc_cfg, "reference_compile"):
|
| 182 |
+
enc_cfg.reference_compile = False
|
| 183 |
+
enc_cfg.save_pretrained(str(out / "encoder"))
|
| 184 |
+
tok.save_pretrained(str(out / "tokenizer"))
|
| 185 |
+
sd = clean_state_dict(model.state_dict())
|
| 186 |
+
fc = dict(model.fcfg)
|
| 187 |
+
fc["temperature"] = model.temperature.detach().float().cpu().tolist()
|
| 188 |
+
fc["weights"] = INT8_FILE if int8 else FP16_FILE
|
| 189 |
+
if int8:
|
| 190 |
+
save_file(quantize_int8(sd), str(out / INT8_FILE), metadata={"format": "falcondec-int8"})
|
| 191 |
+
else:
|
| 192 |
+
save_file({k: (v.half() if v.is_floating_point() else v) for k, v in sd.items()},
|
| 193 |
+
str(out / FP16_FILE), metadata={"format": "falcondec-fp16"})
|
| 194 |
+
(out / CONFIG_FILE).write_text(json.dumps(fc, indent=2, default=str), encoding="utf-8")
|
| 195 |
+
try:
|
| 196 |
+
shutil.copy(__file__, out / "falcondec_modeling.py")
|
| 197 |
+
except Exception:
|
| 198 |
+
pass
|
| 199 |
+
for name, content in (extra_files or {}).items():
|
| 200 |
+
(out / name).write_text(content, encoding="utf-8")
|
| 201 |
+
return out
|
| 202 |
+
|
| 203 |
+
|
| 204 |
+
def load_falcondec(path, device=None, dtype=None, attn_implementation="sdpa"):
|
| 205 |
+
"""Load a FalconDec directory (fp16 or int8) or Hub repo. Returns (model, tokenizer)."""
|
| 206 |
+
from safetensors.torch import load_file
|
| 207 |
+
from transformers import AutoConfig, AutoModel, AutoTokenizer
|
| 208 |
+
p = Path(path)
|
| 209 |
+
if not p.exists():
|
| 210 |
+
from huggingface_hub import snapshot_download
|
| 211 |
+
p = Path(snapshot_download(str(path)))
|
| 212 |
+
fc = json.loads((p / CONFIG_FILE).read_text(encoding="utf-8"))
|
| 213 |
+
ecfg = AutoConfig.from_pretrained(str(p / "encoder"))
|
| 214 |
+
if hasattr(ecfg, "reference_compile"):
|
| 215 |
+
ecfg.reference_compile = False
|
| 216 |
+
try:
|
| 217 |
+
enc = AutoModel.from_config(ecfg, attn_implementation=attn_implementation)
|
| 218 |
+
except Exception:
|
| 219 |
+
enc = AutoModel.from_config(ecfg)
|
| 220 |
+
model = FalconDec(enc, fc)
|
| 221 |
+
wf = p / fc.get("weights", FP16_FILE)
|
| 222 |
+
if not wf.exists():
|
| 223 |
+
wf = p / (INT8_FILE if (p / INT8_FILE).exists() else FP16_FILE)
|
| 224 |
+
sd = load_file(str(wf))
|
| 225 |
+
if wf.name == INT8_FILE:
|
| 226 |
+
sd = dequantize_int8(sd)
|
| 227 |
+
missing, unexpected = model.load_state_dict(sd, strict=False)
|
| 228 |
+
if missing or unexpected:
|
| 229 |
+
print(f"[FalconDec] load warning: missing={list(missing)[:5]} unexpected={list(unexpected)[:5]}")
|
| 230 |
+
tok = AutoTokenizer.from_pretrained(str(p / "tokenizer"))
|
| 231 |
+
dev = torch.device(device) if device is not None else torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
| 232 |
+
model.to(dev)
|
| 233 |
+
if dtype is not None:
|
| 234 |
+
model.to(dtype)
|
| 235 |
+
model.eval()
|
| 236 |
+
return model, tok
|
| 237 |
+
|
| 238 |
+
|
| 239 |
+
# ------------------------------------------------------------------ inference
|
| 240 |
+
def _amp(model):
|
| 241 |
+
if model.device.type == "cuda" and next(model.parameters()).dtype == torch.float32:
|
| 242 |
+
dt = torch.bfloat16 if torch.cuda.is_bf16_supported() else torch.float16
|
| 243 |
+
return torch.autocast("cuda", dtype=dt)
|
| 244 |
+
return contextlib.nullcontext()
|
| 245 |
+
|
| 246 |
+
|
| 247 |
+
def _state_text(state):
|
| 248 |
+
if state is None:
|
| 249 |
+
return ""
|
| 250 |
+
return state if isinstance(state, str) else json.dumps(state, ensure_ascii=False)
|
| 251 |
+
|
| 252 |
+
|
| 253 |
+
@torch.no_grad()
|
| 254 |
+
def score_items(model, tok, items, batch_size=32):
|
| 255 |
+
"""items: [{"state", "question", "options", "type"?, "option_tokens"?, "seq_len"?}] -> list of prob arrays."""
|
| 256 |
+
fc = model.fcfg
|
| 257 |
+
M = int(fc.get("max_opts_single_pass", 96))
|
| 258 |
+
out = [None] * len(items)
|
| 259 |
+
small = [i for i, it in enumerate(items) if len(it["options"]) <= M]
|
| 260 |
+
for i, it in enumerate(items):
|
| 261 |
+
if len(it["options"]) > M:
|
| 262 |
+
out[i] = _score_large(model, tok, it, batch_size)
|
| 263 |
+
if not small:
|
| 264 |
+
return out
|
| 265 |
+
sp = fc["special"]
|
| 266 |
+
prepared = {}
|
| 267 |
+
for i in small:
|
| 268 |
+
it = items[i]
|
| 269 |
+
if len(it["options"]) < 2:
|
| 270 |
+
raise ValueError("Every question needs at least two options.")
|
| 271 |
+
budget = int(it.get("option_tokens", fc["max_tok_per_opt"]))
|
| 272 |
+
q_ids = tok(it.get("question", ""), add_special_tokens=False, truncation=True, max_length=96)["input_ids"]
|
| 273 |
+
o_ids = tok([str(o) for o in it["options"]], add_special_tokens=False, truncation=True,
|
| 274 |
+
max_length=budget)["input_ids"]
|
| 275 |
+
s = _state_text(it.get("state"))
|
| 276 |
+
s_ids = tok(s, add_special_tokens=False, truncation=True,
|
| 277 |
+
max_length=int(fc["long_max_len"]))["input_ids"] if s else []
|
| 278 |
+
ids, mk = assemble(q_ids, o_ids, s_ids, sp, int(it.get("seq_len", fc["max_len"])), fc["head_max_len"],
|
| 279 |
+
budget, fc["long_max_len"], fc["long_opts_threshold"])
|
| 280 |
+
prepared[i] = (ids, mk, QTYPES[it.get("type", "choice")])
|
| 281 |
+
order = sorted(small, key=lambda i: len(prepared[i][0]))
|
| 282 |
+
model.eval()
|
| 283 |
+
for b0 in range(0, len(order), batch_size):
|
| 284 |
+
idxs = order[b0: b0 + batch_size]
|
| 285 |
+
batch = collate_features([prepared[i] for i in idxs], sp["pad"], model.device)
|
| 286 |
+
with _amp(model):
|
| 287 |
+
logits = model(**batch)
|
| 288 |
+
for j, i in enumerate(idxs):
|
| 289 |
+
n = len(prepared[i][1])
|
| 290 |
+
T = model.temperature[prepared[i][2], n_bucket(n)].float().clamp_min(1e-3)
|
| 291 |
+
out[i] = torch.softmax(logits[j, :n].float() / T, -1).cpu().numpy()
|
| 292 |
+
return out
|
| 293 |
+
|
| 294 |
+
|
| 295 |
+
def _score_large(model, tok, it, batch_size):
|
| 296 |
+
"""Tournament for > max_opts_single_pass options: keep the best of each chunk, then one final pass."""
|
| 297 |
+
M = int(model.fcfg.get("max_opts_single_pass", 96))
|
| 298 |
+
opts = list(it["options"])
|
| 299 |
+
cand = list(range(len(opts)))
|
| 300 |
+
while len(cand) > M:
|
| 301 |
+
groups = [cand[i: i + M] for i in range(0, len(cand), M)]
|
| 302 |
+
keep = max(1, M // len(groups))
|
| 303 |
+
ps = score_items(model, tok, [dict(it, options=[opts[c] for c in g]) for g in groups], batch_size)
|
| 304 |
+
cand = [g[t] for g, p in zip(groups, ps) for t in np.argsort(-p)[:keep]]
|
| 305 |
+
p = score_items(model, tok, [dict(it, options=[opts[c] for c in cand])], batch_size)[0]
|
| 306 |
+
full = np.zeros(len(opts), dtype=np.float32)
|
| 307 |
+
full[cand] = p
|
| 308 |
+
return full
|
| 309 |
+
|
| 310 |
+
|
| 311 |
+
def _normalize_question(q):
|
| 312 |
+
qtype = q.get("type", "choice")
|
| 313 |
+
text = q.get("question") or q.get("instructions") or ""
|
| 314 |
+
if qtype == "noul":
|
| 315 |
+
lab = q.get("labels") or {}
|
| 316 |
+
return qtype, text, [True, False], [str(lab.get("true", "Yes")), str(lab.get("false", "No"))]
|
| 317 |
+
crit = q.get("criteria", q.get("options"))
|
| 318 |
+
if qtype == "score":
|
| 319 |
+
opts = [str(c) for c in crit]
|
| 320 |
+
return qtype, text, list(range(len(opts))), opts
|
| 321 |
+
if isinstance(crit, dict):
|
| 322 |
+
return qtype, text, list(crit), [f"{k}: {v}" if v else str(k) for k, v in crit.items()]
|
| 323 |
+
return qtype, text, list(crit), [str(c) for c in crit]
|
| 324 |
+
|
| 325 |
+
|
| 326 |
+
@torch.no_grad()
|
| 327 |
+
def decide(model, tok, state, questions, defer_threshold=None, batch_size=32):
|
| 328 |
+
"""Answer typed questions about one state.
|
| 329 |
+
|
| 330 |
+
questions: list of dicts, or a Jev/Laya-style dict {key: question}. Each question:
|
| 331 |
+
{"type": "choice"|"noul"|"score", "question"/"instructions": str,
|
| 332 |
+
"options": [...] or "criteria": {key: description} / [levels], "option_tokens"?: int}
|
| 333 |
+
"""
|
| 334 |
+
if isinstance(questions, dict):
|
| 335 |
+
questions = [dict(q, key=k) for k, q in questions.items()]
|
| 336 |
+
items, metas = [], []
|
| 337 |
+
for q in questions:
|
| 338 |
+
qtype, text, keys, opts = _normalize_question(q)
|
| 339 |
+
item = {"state": state, "question": text, "options": opts, "type": qtype}
|
| 340 |
+
for extra in ("option_tokens", "seq_len"):
|
| 341 |
+
if extra in q:
|
| 342 |
+
item[extra] = q[extra]
|
| 343 |
+
items.append(item)
|
| 344 |
+
metas.append((q, qtype, keys, opts))
|
| 345 |
+
probs = score_items(model, tok, items, batch_size)
|
| 346 |
+
thr = model.fcfg.get("defer_threshold", 0.0) if defer_threshold is None else defer_threshold
|
| 347 |
+
results = []
|
| 348 |
+
for (q, qtype, keys, opts), p in zip(metas, probs):
|
| 349 |
+
i = int(np.argmax(p))
|
| 350 |
+
r = {"key": q.get("key"), "type": qtype, "question": items[len(results)]["question"],
|
| 351 |
+
"choice": keys[i], "choice_text": opts[i], "confidence": float(p[i]),
|
| 352 |
+
"probs": {str(k): float(v) for k, v in zip(keys, p)}, "defer": float(p[i]) < thr}
|
| 353 |
+
if qtype == "noul":
|
| 354 |
+
r["p_true"] = float(p[0])
|
| 355 |
+
if qtype == "score":
|
| 356 |
+
r["expected_level"] = float(np.dot(p, np.arange(len(p))))
|
| 357 |
+
results.append(r)
|
| 358 |
+
return {"results": results, "answers": {r["key"]: r for r in results if r["key"] is not None}}
|
falcondec_report.json
ADDED
|
@@ -0,0 +1,1802 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"version": "1.0.0",
|
| 3 |
+
"notebook_version": "V3.7",
|
| 4 |
+
"mode": "scratch",
|
| 5 |
+
"parent_version": null,
|
| 6 |
+
"lineage": [
|
| 7 |
+
"jhu-clsp/ettin-encoder-150m"
|
| 8 |
+
],
|
| 9 |
+
"config": {
|
| 10 |
+
"MODE": "scratch",
|
| 11 |
+
"FINETUNE_FROM": "latest",
|
| 12 |
+
"MODEL_NAME": "LightDec_V2_Long",
|
| 13 |
+
"BACKBONE": "jhu-clsp/ettin-encoder-150m",
|
| 14 |
+
"BACKBONE_INIT": "pretrained",
|
| 15 |
+
"PRESET": "long",
|
| 16 |
+
"CUSTOM_DATA_JSONL": "",
|
| 17 |
+
"CUSTOM_WEIGHT": 3.0,
|
| 18 |
+
"INCLUDE_BUILDERS": [],
|
| 19 |
+
"EXCLUDE_BUILDERS": [
|
| 20 |
+
"mind2web"
|
| 21 |
+
],
|
| 22 |
+
"TEV1_DIR": "./tev1",
|
| 23 |
+
"TEV1_AUTOBUILD": true,
|
| 24 |
+
"TEV1_POLICY_CAP": 13500,
|
| 25 |
+
"TEV1_ROUTING_CAP": 6000,
|
| 26 |
+
"TEV1_RESEARCH_CAP": 1000,
|
| 27 |
+
"TRIAGE_EMAILS": 200,
|
| 28 |
+
"TEV1_RETRIES": 3,
|
| 29 |
+
"TEV1_STEP_TIMEOUT_MIN": 180,
|
| 30 |
+
"TEV1_STALL_MIN": 15,
|
| 31 |
+
"RETRY_FAILED_BUILDERS": true,
|
| 32 |
+
"LONG_TASK_TRAIN": 1500,
|
| 33 |
+
"MAX_LEN": 2048,
|
| 34 |
+
"LONG_MAX_LEN": 2048,
|
| 35 |
+
"LONG_OPTS_THRESHOLD": 24,
|
| 36 |
+
"HEAD_MAX_LEN": 192,
|
| 37 |
+
"MAX_TOK_PER_OPT": 24,
|
| 38 |
+
"MAX_OPTS_SINGLE_PASS": 96,
|
| 39 |
+
"INTERACT_LAYERS": 2,
|
| 40 |
+
"INTERACT_HEADS": 8,
|
| 41 |
+
"BATCH_SIZE": 128,
|
| 42 |
+
"TOKENS_PER_BATCH": 65536,
|
| 43 |
+
"GRAD_ACCUM": 1,
|
| 44 |
+
"LR_ENCODER": 8e-05,
|
| 45 |
+
"LR_HEAD": 0.0006,
|
| 46 |
+
"LLRD": 0.9,
|
| 47 |
+
"WEIGHT_DECAY": 0.01,
|
| 48 |
+
"WARMUP_FRAC": 0.06,
|
| 49 |
+
"EPOCHS": 0,
|
| 50 |
+
"EMA_DECAY": 0.999,
|
| 51 |
+
"GRAD_CLIP": 1.0,
|
| 52 |
+
"FINETUNE_LR_SCALE": 0.5,
|
| 53 |
+
"SPHERICAL_W": 0.5,
|
| 54 |
+
"BRIER_W": 0.0,
|
| 55 |
+
"RPS_W": 1.0,
|
| 56 |
+
"TASK_SAMPLING_ALPHA": 0.5,
|
| 57 |
+
"NOTA_PROB": 0.08,
|
| 58 |
+
"RLCD_EPOCHS": 0.0,
|
| 59 |
+
"RLCD_SAMPLES": 4,
|
| 60 |
+
"RLCD_SIGMA": 0.3,
|
| 61 |
+
"RLCD_CE_ANCHOR": 0.5,
|
| 62 |
+
"DEVICE": "cuda",
|
| 63 |
+
"SEED": 42,
|
| 64 |
+
"NUM_WORKERS": 8,
|
| 65 |
+
"USE_COMPILE": true,
|
| 66 |
+
"ATTN_IMPL": "auto",
|
| 67 |
+
"OUTPUT_ROOT": "./falcondec_runs",
|
| 68 |
+
"EXPORT_INT8": true,
|
| 69 |
+
"SIZE_TARGET_MB": 250,
|
| 70 |
+
"DEFER_THRESHOLD": 0.7,
|
| 71 |
+
"EVAL_BASELINE_PROOF_V2": true,
|
| 72 |
+
"BASELINE_MAX_PER_TASK": 120,
|
| 73 |
+
"RUN_CPU_BENCH": true,
|
| 74 |
+
"PUSH_TO_HUB": false,
|
| 75 |
+
"HUB_REPO": ""
|
| 76 |
+
},
|
| 77 |
+
"epochs": 10,
|
| 78 |
+
"best_epoch": 4,
|
| 79 |
+
"rlcd_kept": false,
|
| 80 |
+
"history": [
|
| 81 |
+
{
|
| 82 |
+
"epoch": 1,
|
| 83 |
+
"train_loss": 0.7058714814313827,
|
| 84 |
+
"train_acc": 0.7342094450703411,
|
| 85 |
+
"val_macro": 0.8188568656508199,
|
| 86 |
+
"val_micro": 0.820702479338843,
|
| 87 |
+
"val_nll": 0.4076740937270036
|
| 88 |
+
},
|
| 89 |
+
{
|
| 90 |
+
"epoch": 2,
|
| 91 |
+
"train_loss": 0.431469622218984,
|
| 92 |
+
"train_acc": 0.8473141073247583,
|
| 93 |
+
"val_macro": 0.8450872188411825,
|
| 94 |
+
"val_micro": 0.8444628099173553,
|
| 95 |
+
"val_nll": 0.3623910443100994
|
| 96 |
+
},
|
| 97 |
+
{
|
| 98 |
+
"epoch": 3,
|
| 99 |
+
"train_loss": 0.34276796075090776,
|
| 100 |
+
"train_acc": 0.8791940521680108,
|
| 101 |
+
"val_macro": 0.8501904675940947,
|
| 102 |
+
"val_micro": 0.8499173553719008,
|
| 103 |
+
"val_nll": 0.39227130460368825
|
| 104 |
+
},
|
| 105 |
+
{
|
| 106 |
+
"epoch": 4,
|
| 107 |
+
"train_loss": 0.2773282952523419,
|
| 108 |
+
"train_acc": 0.9039863196679513,
|
| 109 |
+
"val_macro": 0.8519172452047367,
|
| 110 |
+
"val_micro": 0.8519834710743802,
|
| 111 |
+
"val_nll": 0.4529550437184426
|
| 112 |
+
},
|
| 113 |
+
{
|
| 114 |
+
"epoch": 5,
|
| 115 |
+
"train_loss": 0.22030453415052323,
|
| 116 |
+
"train_acc": 0.925851485588573,
|
| 117 |
+
"val_macro": 0.8488092001804082,
|
| 118 |
+
"val_micro": 0.85,
|
| 119 |
+
"val_nll": 0.5369876145373608
|
| 120 |
+
},
|
| 121 |
+
{
|
| 122 |
+
"epoch": 6,
|
| 123 |
+
"train_loss": 0.16953414903032543,
|
| 124 |
+
"train_acc": 0.9449156076226879,
|
| 125 |
+
"val_macro": 0.8474279008050457,
|
| 126 |
+
"val_micro": 0.8488842975206612,
|
| 127 |
+
"val_nll": 0.6705887193589142
|
| 128 |
+
},
|
| 129 |
+
{
|
| 130 |
+
"epoch": 7,
|
| 131 |
+
"train_loss": 0.126976966611751,
|
| 132 |
+
"train_acc": 0.9602833527848939,
|
| 133 |
+
"val_macro": 0.8453322694829587,
|
| 134 |
+
"val_micro": 0.8476446280991735,
|
| 135 |
+
"val_nll": 0.8430423012207805
|
| 136 |
+
},
|
| 137 |
+
{
|
| 138 |
+
"epoch": 8,
|
| 139 |
+
"train_loss": 0.09470793687766035,
|
| 140 |
+
"train_acc": 0.9714013975720073,
|
| 141 |
+
"val_macro": 0.8451769009269958,
|
| 142 |
+
"val_micro": 0.8467355371900827,
|
| 143 |
+
"val_nll": 1.0336680755654943
|
| 144 |
+
},
|
| 145 |
+
{
|
| 146 |
+
"epoch": 9,
|
| 147 |
+
"train_loss": 0.0710737590306488,
|
| 148 |
+
"train_acc": 0.9794225926173407,
|
| 149 |
+
"val_macro": 0.8435159403610613,
|
| 150 |
+
"val_micro": 0.8458677685950413,
|
| 151 |
+
"val_nll": 1.210751206553627
|
| 152 |
+
},
|
| 153 |
+
{
|
| 154 |
+
"epoch": 10,
|
| 155 |
+
"train_loss": 0.05665995772408678,
|
| 156 |
+
"train_acc": 0.9843832093140154,
|
| 157 |
+
"val_macro": 0.844272704258198,
|
| 158 |
+
"val_micro": 0.8474380165289256,
|
| 159 |
+
"val_nll": 1.41872131482475
|
| 160 |
+
}
|
| 161 |
+
],
|
| 162 |
+
"train_minutes": 291.67019440333047,
|
| 163 |
+
"temperatures": [
|
| 164 |
+
[
|
| 165 |
+
2.37591814994812,
|
| 166 |
+
2.0426506996154785,
|
| 167 |
+
1.6455482244491577,
|
| 168 |
+
1.8417285680770874
|
| 169 |
+
],
|
| 170 |
+
[
|
| 171 |
+
2.4885058403015137,
|
| 172 |
+
2.4885058403015137,
|
| 173 |
+
2.4885058403015137,
|
| 174 |
+
2.4885058403015137
|
| 175 |
+
],
|
| 176 |
+
[
|
| 177 |
+
2.2925639152526855,
|
| 178 |
+
2.2925639152526855,
|
| 179 |
+
2.2925639152526855,
|
| 180 |
+
2.2925639152526855
|
| 181 |
+
]
|
| 182 |
+
],
|
| 183 |
+
"data_counts": {
|
| 184 |
+
"train": 1023814,
|
| 185 |
+
"validation": 24200,
|
| 186 |
+
"test": 31990
|
| 187 |
+
},
|
| 188 |
+
"skipped_builders": [],
|
| 189 |
+
"test_overall": {
|
| 190 |
+
"n": 31990.0,
|
| 191 |
+
"micro_acc": 0.784057517974367,
|
| 192 |
+
"macro_acc": 0.7887006006875437,
|
| 193 |
+
"nll": 0.5630830966790886,
|
| 194 |
+
"brier": 0.2954194093899411,
|
| 195 |
+
"ece": 0.02856345933167103,
|
| 196 |
+
"aurc": 0.06778106632931055,
|
| 197 |
+
"score_mae": 0.5060575584353655,
|
| 198 |
+
"coverage@0.7": 0.6797436698968428,
|
| 199 |
+
"acc_on_covered@0.7": 0.9059094044607956
|
| 200 |
+
},
|
| 201 |
+
"test_per_domain": {
|
| 202 |
+
"agentic": 0.821117397177404,
|
| 203 |
+
"classification": 0.6115,
|
| 204 |
+
"code": 0.9108344674425007,
|
| 205 |
+
"guardrails": 0.7984073149310548,
|
| 206 |
+
"intents": 0.8505,
|
| 207 |
+
"long_context": 1.0,
|
| 208 |
+
"policy": 0.8061818181818182,
|
| 209 |
+
"reasoning": 0.7180877154615182,
|
| 210 |
+
"support": 0.9815181518151815,
|
| 211 |
+
"tev1_benchmark": 0.6541666666666667,
|
| 212 |
+
"workflows": 0.773
|
| 213 |
+
},
|
| 214 |
+
"test_per_task": [
|
| 215 |
+
{
|
| 216 |
+
"task": "ag_news/topic",
|
| 217 |
+
"n": 500,
|
| 218 |
+
"acc": 0.9,
|
| 219 |
+
"chance": 0.25,
|
| 220 |
+
"nll": 0.2674782864307199,
|
| 221 |
+
"brier": 0.14296993909824685,
|
| 222 |
+
"heldout": false,
|
| 223 |
+
"domain": "classification",
|
| 224 |
+
"ece": 0.05165432274341582,
|
| 225 |
+
"proof_v2 (card)": NaN,
|
| 226 |
+
"\u0394 vs proof_v2": NaN
|
| 227 |
+
},
|
| 228 |
+
{
|
| 229 |
+
"task": "agentharm/refuse",
|
| 230 |
+
"n": 416,
|
| 231 |
+
"acc": 0.6995192307692307,
|
| 232 |
+
"chance": 0.5,
|
| 233 |
+
"nll": 0.6008703433729422,
|
| 234 |
+
"brier": 0.4076762512706606,
|
| 235 |
+
"heldout": true,
|
| 236 |
+
"domain": "guardrails",
|
| 237 |
+
"ece": 0.08916855761064935,
|
| 238 |
+
"proof_v2 (card)": NaN,
|
| 239 |
+
"\u0394 vs proof_v2": NaN
|
| 240 |
+
},
|
| 241 |
+
{
|
| 242 |
+
"task": "agenttrek/finish_now",
|
| 243 |
+
"n": 151,
|
| 244 |
+
"acc": 0.7880794701986755,
|
| 245 |
+
"chance": 0.5,
|
| 246 |
+
"nll": 0.445978463758558,
|
| 247 |
+
"brier": 0.30066953632150406,
|
| 248 |
+
"heldout": false,
|
| 249 |
+
"domain": "agentic",
|
| 250 |
+
"ece": 0.07512128787324918,
|
| 251 |
+
"proof_v2 (card)": NaN,
|
| 252 |
+
"\u0394 vs proof_v2": NaN
|
| 253 |
+
},
|
| 254 |
+
{
|
| 255 |
+
"task": "agenttrek/next_action_type",
|
| 256 |
+
"n": 487,
|
| 257 |
+
"acc": 0.864476386036961,
|
| 258 |
+
"chance": 0.19054952576513148,
|
| 259 |
+
"nll": 0.3929848723152591,
|
| 260 |
+
"brier": 0.2184262054212701,
|
| 261 |
+
"heldout": false,
|
| 262 |
+
"domain": "agentic",
|
| 263 |
+
"ece": 0.09202881025827395,
|
| 264 |
+
"proof_v2 (card)": NaN,
|
| 265 |
+
"\u0394 vs proof_v2": NaN
|
| 266 |
+
},
|
| 267 |
+
{
|
| 268 |
+
"task": "anli/nli",
|
| 269 |
+
"n": 498,
|
| 270 |
+
"acc": 0.4859437751004016,
|
| 271 |
+
"chance": 0.33333333333333326,
|
| 272 |
+
"nll": 1.069221544292677,
|
| 273 |
+
"brier": 0.6544603492386665,
|
| 274 |
+
"heldout": false,
|
| 275 |
+
"domain": "reasoning",
|
| 276 |
+
"ece": 0.14589204983299517,
|
| 277 |
+
"proof_v2 (card)": NaN,
|
| 278 |
+
"\u0394 vs proof_v2": NaN
|
| 279 |
+
},
|
| 280 |
+
{
|
| 281 |
+
"task": "aqua_rat/math",
|
| 282 |
+
"n": 247,
|
| 283 |
+
"acc": 0.340080971659919,
|
| 284 |
+
"chance": 0.2,
|
| 285 |
+
"nll": 1.54525283087603,
|
| 286 |
+
"brier": 0.7739737508966391,
|
| 287 |
+
"heldout": false,
|
| 288 |
+
"domain": "reasoning",
|
| 289 |
+
"ece": 0.080350092668765,
|
| 290 |
+
"proof_v2 (card)": NaN,
|
| 291 |
+
"\u0394 vs proof_v2": NaN
|
| 292 |
+
},
|
| 293 |
+
{
|
| 294 |
+
"task": "arc_challenge/mcq",
|
| 295 |
+
"n": 500,
|
| 296 |
+
"acc": 0.49,
|
| 297 |
+
"chance": 0.25006666666666666,
|
| 298 |
+
"nll": 1.3021090898208785,
|
| 299 |
+
"brier": 0.6752390805252713,
|
| 300 |
+
"heldout": true,
|
| 301 |
+
"domain": "reasoning",
|
| 302 |
+
"ece": 0.1680490539073944,
|
| 303 |
+
"proof_v2 (card)": 0.308,
|
| 304 |
+
"\u0394 vs proof_v2": 0.182
|
| 305 |
+
},
|
| 306 |
+
{
|
| 307 |
+
"task": "arc_easy/mcq",
|
| 308 |
+
"n": 500,
|
| 309 |
+
"acc": 0.612,
|
| 310 |
+
"chance": 0.24980000000000002,
|
| 311 |
+
"nll": 0.9830388274090365,
|
| 312 |
+
"brier": 0.5264415342806388,
|
| 313 |
+
"heldout": true,
|
| 314 |
+
"domain": "reasoning",
|
| 315 |
+
"ece": 0.1215095258355141,
|
| 316 |
+
"proof_v2 (card)": 0.425,
|
| 317 |
+
"\u0394 vs proof_v2": 0.187
|
| 318 |
+
},
|
| 319 |
+
{
|
| 320 |
+
"task": "banking77/intent",
|
| 321 |
+
"n": 500,
|
| 322 |
+
"acc": 0.922,
|
| 323 |
+
"chance": 0.3214666666666666,
|
| 324 |
+
"nll": 0.23128961177584006,
|
| 325 |
+
"brier": 0.11756106243365176,
|
| 326 |
+
"heldout": true,
|
| 327 |
+
"domain": "intents",
|
| 328 |
+
"ece": 0.037492330908775316,
|
| 329 |
+
"proof_v2 (card)": 0.883,
|
| 330 |
+
"\u0394 vs proof_v2": 0.039000000000000035
|
| 331 |
+
},
|
| 332 |
+
{
|
| 333 |
+
"task": "banking77/intent_77",
|
| 334 |
+
"n": 300,
|
| 335 |
+
"acc": 0.58,
|
| 336 |
+
"chance": 0.012987012987012986,
|
| 337 |
+
"nll": 2.159930980304489,
|
| 338 |
+
"brier": 0.6431857202493357,
|
| 339 |
+
"heldout": true,
|
| 340 |
+
"domain": "intents",
|
| 341 |
+
"ece": 0.24079640249411272,
|
| 342 |
+
"proof_v2 (card)": NaN,
|
| 343 |
+
"\u0394 vs proof_v2": NaN
|
| 344 |
+
},
|
| 345 |
+
{
|
| 346 |
+
"task": "bigclonebench/clone",
|
| 347 |
+
"n": 500,
|
| 348 |
+
"acc": 0.962,
|
| 349 |
+
"chance": 0.5,
|
| 350 |
+
"nll": 0.1312969167594565,
|
| 351 |
+
"brier": 0.06444151702850223,
|
| 352 |
+
"heldout": false,
|
| 353 |
+
"domain": "code",
|
| 354 |
+
"ece": 0.03135248208045959,
|
| 355 |
+
"proof_v2 (card)": 0.383,
|
| 356 |
+
"\u0394 vs proof_v2": 0.579
|
| 357 |
+
},
|
| 358 |
+
{
|
| 359 |
+
"task": "bitext/category",
|
| 360 |
+
"n": 500,
|
| 361 |
+
"acc": 1.0,
|
| 362 |
+
"chance": 0.16128124098124094,
|
| 363 |
+
"nll": 0.0013673773490934594,
|
| 364 |
+
"brier": 0.0003828941821233322,
|
| 365 |
+
"heldout": false,
|
| 366 |
+
"domain": "support",
|
| 367 |
+
"ece": 0.001247308015823329,
|
| 368 |
+
"proof_v2 (card)": NaN,
|
| 369 |
+
"\u0394 vs proof_v2": NaN
|
| 370 |
+
},
|
| 371 |
+
{
|
| 372 |
+
"task": "bitext/route",
|
| 373 |
+
"n": 500,
|
| 374 |
+
"acc": 1.0,
|
| 375 |
+
"chance": 0.19016313131313134,
|
| 376 |
+
"nll": 0.00031508885690954,
|
| 377 |
+
"brier": 4.903548307607296e-06,
|
| 378 |
+
"heldout": false,
|
| 379 |
+
"domain": "support",
|
| 380 |
+
"ece": 0.0003137041330337764,
|
| 381 |
+
"proof_v2 (card)": 0.958,
|
| 382 |
+
"\u0394 vs proof_v2": 0.04200000000000004
|
| 383 |
+
},
|
| 384 |
+
{
|
| 385 |
+
"task": "boolq/yes_no",
|
| 386 |
+
"n": 500,
|
| 387 |
+
"acc": 0.822,
|
| 388 |
+
"chance": 0.5,
|
| 389 |
+
"nll": 0.4842173567947466,
|
| 390 |
+
"brier": 0.2840138305161832,
|
| 391 |
+
"heldout": false,
|
| 392 |
+
"domain": "reasoning",
|
| 393 |
+
"ece": 0.0795972980260849,
|
| 394 |
+
"proof_v2 (card)": 0.717,
|
| 395 |
+
"\u0394 vs proof_v2": 0.10499999999999998
|
| 396 |
+
},
|
| 397 |
+
{
|
| 398 |
+
"task": "civil_comments/toxic",
|
| 399 |
+
"n": 500,
|
| 400 |
+
"acc": 0.926,
|
| 401 |
+
"chance": 0.5,
|
| 402 |
+
"nll": 0.2234244500696659,
|
| 403 |
+
"brier": 0.1209681824839501,
|
| 404 |
+
"heldout": false,
|
| 405 |
+
"domain": "guardrails",
|
| 406 |
+
"ece": 0.05123518884181974,
|
| 407 |
+
"proof_v2 (card)": NaN,
|
| 408 |
+
"\u0394 vs proof_v2": NaN
|
| 409 |
+
},
|
| 410 |
+
{
|
| 411 |
+
"task": "clinc150/intent",
|
| 412 |
+
"n": 500,
|
| 413 |
+
"acc": 0.97,
|
| 414 |
+
"chance": 0.12404474984829318,
|
| 415 |
+
"nll": 0.09906445816022005,
|
| 416 |
+
"brier": 0.04996487012600131,
|
| 417 |
+
"heldout": false,
|
| 418 |
+
"domain": "intents",
|
| 419 |
+
"ece": 0.014944501101970714,
|
| 420 |
+
"proof_v2 (card)": 0.85,
|
| 421 |
+
"\u0394 vs proof_v2": 0.12
|
| 422 |
+
},
|
| 423 |
+
{
|
| 424 |
+
"task": "codexglue/code_to_doc",
|
| 425 |
+
"n": 504,
|
| 426 |
+
"acc": 0.9900793650793651,
|
| 427 |
+
"chance": 0.26233465608465617,
|
| 428 |
+
"nll": 0.032474144255576765,
|
| 429 |
+
"brier": 0.016158895201785088,
|
| 430 |
+
"heldout": false,
|
| 431 |
+
"domain": "code",
|
| 432 |
+
"ece": 0.009654374349685009,
|
| 433 |
+
"proof_v2 (card)": 0.969,
|
| 434 |
+
"\u0394 vs proof_v2": 0.021079365079365142
|
| 435 |
+
},
|
| 436 |
+
{
|
| 437 |
+
"task": "codexglue/doc_to_code",
|
| 438 |
+
"n": 504,
|
| 439 |
+
"acc": 0.9761904761904762,
|
| 440 |
+
"chance": 0.2786044973544973,
|
| 441 |
+
"nll": 0.05970113779317181,
|
| 442 |
+
"brier": 0.03072688989396561,
|
| 443 |
+
"heldout": false,
|
| 444 |
+
"domain": "code",
|
| 445 |
+
"ece": 0.012547755761752008,
|
| 446 |
+
"proof_v2 (card)": 0.961,
|
| 447 |
+
"\u0394 vs proof_v2": 0.015190476190476199
|
| 448 |
+
},
|
| 449 |
+
{
|
| 450 |
+
"task": "codexglue/func_name",
|
| 451 |
+
"n": 467,
|
| 452 |
+
"acc": 0.9550321199143469,
|
| 453 |
+
"chance": 0.25499643112062814,
|
| 454 |
+
"nll": 0.14961532151979784,
|
| 455 |
+
"brier": 0.07053876912275295,
|
| 456 |
+
"heldout": false,
|
| 457 |
+
"domain": "code",
|
| 458 |
+
"ece": 0.03948740715388316,
|
| 459 |
+
"proof_v2 (card)": 0.901,
|
| 460 |
+
"\u0394 vs proof_v2": 0.054032119914346866
|
| 461 |
+
},
|
| 462 |
+
{
|
| 463 |
+
"task": "codexglue/lang_id",
|
| 464 |
+
"n": 504,
|
| 465 |
+
"acc": 1.0,
|
| 466 |
+
"chance": 0.2354166666666667,
|
| 467 |
+
"nll": 0.0014352430686462487,
|
| 468 |
+
"brier": 1.2864369183890018e-05,
|
| 469 |
+
"heldout": false,
|
| 470 |
+
"domain": "code",
|
| 471 |
+
"ece": 0.0014309174721203188,
|
| 472 |
+
"proof_v2 (card)": 0.997,
|
| 473 |
+
"\u0394 vs proof_v2": 0.0030000000000000027
|
| 474 |
+
},
|
| 475 |
+
{
|
| 476 |
+
"task": "commonsense_qa/mcq",
|
| 477 |
+
"n": 493,
|
| 478 |
+
"acc": 0.6430020283975659,
|
| 479 |
+
"chance": 0.19999999999999998,
|
| 480 |
+
"nll": 0.9523430099929484,
|
| 481 |
+
"brier": 0.49435087368161623,
|
| 482 |
+
"heldout": false,
|
| 483 |
+
"domain": "reasoning",
|
| 484 |
+
"ece": 0.14308575696442233,
|
| 485 |
+
"proof_v2 (card)": 0.442,
|
| 486 |
+
"\u0394 vs proof_v2": 0.20100202839756592
|
| 487 |
+
},
|
| 488 |
+
{
|
| 489 |
+
"task": "counsel/critique_quality",
|
| 490 |
+
"n": 201,
|
| 491 |
+
"acc": 0.6268656716417911,
|
| 492 |
+
"chance": 0.3333333333333333,
|
| 493 |
+
"nll": 1.5466346353110312,
|
| 494 |
+
"brier": 0.6427543071457474,
|
| 495 |
+
"heldout": false,
|
| 496 |
+
"domain": "agentic",
|
| 497 |
+
"ece": 0.27062572442477023,
|
| 498 |
+
"proof_v2 (card)": NaN,
|
| 499 |
+
"\u0394 vs proof_v2": NaN
|
| 500 |
+
},
|
| 501 |
+
{
|
| 502 |
+
"task": "counsel/step_has_error",
|
| 503 |
+
"n": 201,
|
| 504 |
+
"acc": 0.8308457711442786,
|
| 505 |
+
"chance": 0.5,
|
| 506 |
+
"nll": 0.8236887436017318,
|
| 507 |
+
"brier": 0.3201374357369142,
|
| 508 |
+
"heldout": false,
|
| 509 |
+
"domain": "agentic",
|
| 510 |
+
"ece": 0.14827988011326954,
|
| 511 |
+
"proof_v2 (card)": NaN,
|
| 512 |
+
"\u0394 vs proof_v2": NaN
|
| 513 |
+
},
|
| 514 |
+
{
|
| 515 |
+
"task": "devign/vulnerability",
|
| 516 |
+
"n": 500,
|
| 517 |
+
"acc": 0.632,
|
| 518 |
+
"chance": 0.5,
|
| 519 |
+
"nll": 0.6074751503933221,
|
| 520 |
+
"brier": 0.42706214521965163,
|
| 521 |
+
"heldout": false,
|
| 522 |
+
"domain": "code",
|
| 523 |
+
"ece": 0.07018342161178594,
|
| 524 |
+
"proof_v2 (card)": 0.542,
|
| 525 |
+
"\u0394 vs proof_v2": 0.08999999999999997
|
| 526 |
+
},
|
| 527 |
+
{
|
| 528 |
+
"task": "emotion/6way",
|
| 529 |
+
"n": 500,
|
| 530 |
+
"acc": 0.472,
|
| 531 |
+
"chance": 0.16666666666666663,
|
| 532 |
+
"nll": 1.561316128242761,
|
| 533 |
+
"brier": 0.7356062387487318,
|
| 534 |
+
"heldout": true,
|
| 535 |
+
"domain": "classification",
|
| 536 |
+
"ece": 0.21054179659485817,
|
| 537 |
+
"proof_v2 (card)": NaN,
|
| 538 |
+
"\u0394 vs proof_v2": NaN
|
| 539 |
+
},
|
| 540 |
+
{
|
| 541 |
+
"task": "gsm8k/math",
|
| 542 |
+
"n": 500,
|
| 543 |
+
"acc": 0.7,
|
| 544 |
+
"chance": 0.25,
|
| 545 |
+
"nll": 0.6923719964642078,
|
| 546 |
+
"brier": 0.4014884319624292,
|
| 547 |
+
"heldout": false,
|
| 548 |
+
"domain": "reasoning",
|
| 549 |
+
"ece": 0.08541783547401426,
|
| 550 |
+
"proof_v2 (card)": 0.275,
|
| 551 |
+
"\u0394 vs proof_v2": 0.42499999999999993
|
| 552 |
+
},
|
| 553 |
+
{
|
| 554 |
+
"task": "hellaswag/continuation",
|
| 555 |
+
"n": 500,
|
| 556 |
+
"acc": 0.576,
|
| 557 |
+
"chance": 0.25,
|
| 558 |
+
"nll": 0.968790470642969,
|
| 559 |
+
"brier": 0.5288133586177948,
|
| 560 |
+
"heldout": false,
|
| 561 |
+
"domain": "reasoning",
|
| 562 |
+
"ece": 0.07543198531866073,
|
| 563 |
+
"proof_v2 (card)": NaN,
|
| 564 |
+
"\u0394 vs proof_v2": NaN
|
| 565 |
+
},
|
| 566 |
+
{
|
| 567 |
+
"task": "hotpotqa/comparison_yes_no",
|
| 568 |
+
"n": 26,
|
| 569 |
+
"acc": 0.9230769230769231,
|
| 570 |
+
"chance": 0.5,
|
| 571 |
+
"nll": 0.3400710394176153,
|
| 572 |
+
"brier": 0.15484679762361436,
|
| 573 |
+
"heldout": false,
|
| 574 |
+
"domain": "agentic",
|
| 575 |
+
"ece": 0.08232885369887719,
|
| 576 |
+
"proof_v2 (card)": NaN,
|
| 577 |
+
"\u0394 vs proof_v2": NaN
|
| 578 |
+
},
|
| 579 |
+
{
|
| 580 |
+
"task": "hotpotqa/retrieve",
|
| 581 |
+
"n": 497,
|
| 582 |
+
"acc": 0.8933601609657947,
|
| 583 |
+
"chance": 0.16741720162243304,
|
| 584 |
+
"nll": 0.3457089332123268,
|
| 585 |
+
"brier": 0.16216258689991922,
|
| 586 |
+
"heldout": false,
|
| 587 |
+
"domain": "agentic",
|
| 588 |
+
"ece": 0.04268557528854612,
|
| 589 |
+
"proof_v2 (card)": NaN,
|
| 590 |
+
"\u0394 vs proof_v2": NaN
|
| 591 |
+
},
|
| 592 |
+
{
|
| 593 |
+
"task": "humaneval/completion",
|
| 594 |
+
"n": 119,
|
| 595 |
+
"acc": 0.8487394957983193,
|
| 596 |
+
"chance": 0.39355742296918783,
|
| 597 |
+
"nll": 0.46361927195851294,
|
| 598 |
+
"brier": 0.2730045795813753,
|
| 599 |
+
"heldout": true,
|
| 600 |
+
"domain": "code",
|
| 601 |
+
"ece": 0.13989568708323633,
|
| 602 |
+
"proof_v2 (card)": 0.575,
|
| 603 |
+
"\u0394 vs proof_v2": 0.27373949579831935
|
| 604 |
+
},
|
| 605 |
+
{
|
| 606 |
+
"task": "jailbreak/detect",
|
| 607 |
+
"n": 262,
|
| 608 |
+
"acc": 0.9732824427480916,
|
| 609 |
+
"chance": 0.5,
|
| 610 |
+
"nll": 0.08469655378148465,
|
| 611 |
+
"brier": 0.044003931208438284,
|
| 612 |
+
"heldout": false,
|
| 613 |
+
"domain": "guardrails",
|
| 614 |
+
"ece": 0.030764950368240628,
|
| 615 |
+
"proof_v2 (card)": NaN,
|
| 616 |
+
"\u0394 vs proof_v2": NaN
|
| 617 |
+
},
|
| 618 |
+
{
|
| 619 |
+
"task": "long/contract_clause",
|
| 620 |
+
"n": 150,
|
| 621 |
+
"acc": 1.0,
|
| 622 |
+
"chance": 0.19999999999999996,
|
| 623 |
+
"nll": 0.0002789537957808837,
|
| 624 |
+
"brier": 4.274148862877054e-07,
|
| 625 |
+
"heldout": false,
|
| 626 |
+
"domain": "long_context",
|
| 627 |
+
"ece": 0.00027882695198055973,
|
| 628 |
+
"proof_v2 (card)": NaN,
|
| 629 |
+
"\u0394 vs proof_v2": NaN
|
| 630 |
+
},
|
| 631 |
+
{
|
| 632 |
+
"task": "long/email_thread",
|
| 633 |
+
"n": 150,
|
| 634 |
+
"acc": 1.0,
|
| 635 |
+
"chance": 0.25,
|
| 636 |
+
"nll": 0.0003266528345905802,
|
| 637 |
+
"brier": 2.717841913247229e-06,
|
| 638 |
+
"heldout": false,
|
| 639 |
+
"domain": "long_context",
|
| 640 |
+
"ece": 0.0003259424368540209,
|
| 641 |
+
"proof_v2 (card)": NaN,
|
| 642 |
+
"\u0394 vs proof_v2": NaN
|
| 643 |
+
},
|
| 644 |
+
{
|
| 645 |
+
"task": "long/service_log",
|
| 646 |
+
"n": 150,
|
| 647 |
+
"acc": 1.0,
|
| 648 |
+
"chance": 0.19999999999999996,
|
| 649 |
+
"nll": 0.0010035930563268873,
|
| 650 |
+
"brier": 2.0987077842670875e-06,
|
| 651 |
+
"heldout": false,
|
| 652 |
+
"domain": "long_context",
|
| 653 |
+
"ece": 0.0010028688112894146,
|
| 654 |
+
"proof_v2 (card)": NaN,
|
| 655 |
+
"\u0394 vs proof_v2": NaN
|
| 656 |
+
},
|
| 657 |
+
{
|
| 658 |
+
"task": "massive_en/intent",
|
| 659 |
+
"n": 500,
|
| 660 |
+
"acc": 0.93,
|
| 661 |
+
"chance": 0.12958727789085603,
|
| 662 |
+
"nll": 0.20636169074449573,
|
| 663 |
+
"brier": 0.10375277943487987,
|
| 664 |
+
"heldout": false,
|
| 665 |
+
"domain": "intents",
|
| 666 |
+
"ece": 0.0284088468849659,
|
| 667 |
+
"proof_v2 (card)": NaN,
|
| 668 |
+
"\u0394 vs proof_v2": NaN
|
| 669 |
+
},
|
| 670 |
+
{
|
| 671 |
+
"task": "mbpp/bugspot",
|
| 672 |
+
"n": 256,
|
| 673 |
+
"acc": 0.85546875,
|
| 674 |
+
"chance": 0.40559895833333326,
|
| 675 |
+
"nll": 0.3667814989712497,
|
| 676 |
+
"brier": 0.21843909652329055,
|
| 677 |
+
"heldout": false,
|
| 678 |
+
"domain": "code",
|
| 679 |
+
"ece": 0.037241218378767364,
|
| 680 |
+
"proof_v2 (card)": 0.475,
|
| 681 |
+
"\u0394 vs proof_v2": 0.38046875
|
| 682 |
+
},
|
| 683 |
+
{
|
| 684 |
+
"task": "mbpp/solution",
|
| 685 |
+
"n": 500,
|
| 686 |
+
"acc": 0.978,
|
| 687 |
+
"chance": 0.25,
|
| 688 |
+
"nll": 0.05863042902501184,
|
| 689 |
+
"brier": 0.02981698831494921,
|
| 690 |
+
"heldout": false,
|
| 691 |
+
"domain": "code",
|
| 692 |
+
"ece": 0.01960155099630357,
|
| 693 |
+
"proof_v2 (card)": 0.992,
|
| 694 |
+
"\u0394 vs proof_v2": -0.014000000000000012
|
| 695 |
+
},
|
| 696 |
+
{
|
| 697 |
+
"task": "mmlu/mcq",
|
| 698 |
+
"n": 500,
|
| 699 |
+
"acc": 0.39,
|
| 700 |
+
"chance": 0.25,
|
| 701 |
+
"nll": 1.4543649319559335,
|
| 702 |
+
"brier": 0.7702575193705765,
|
| 703 |
+
"heldout": true,
|
| 704 |
+
"domain": "reasoning",
|
| 705 |
+
"ece": 0.15507307499647138,
|
| 706 |
+
"proof_v2 (card)": NaN,
|
| 707 |
+
"\u0394 vs proof_v2": NaN
|
| 708 |
+
},
|
| 709 |
+
{
|
| 710 |
+
"task": "mnli/claim",
|
| 711 |
+
"n": 500,
|
| 712 |
+
"acc": 0.86,
|
| 713 |
+
"chance": 0.33333333333333326,
|
| 714 |
+
"nll": 0.4006310784481466,
|
| 715 |
+
"brier": 0.2131232579488494,
|
| 716 |
+
"heldout": false,
|
| 717 |
+
"domain": "reasoning",
|
| 718 |
+
"ece": 0.081218329668045,
|
| 719 |
+
"proof_v2 (card)": 0.492,
|
| 720 |
+
"\u0394 vs proof_v2": 0.368
|
| 721 |
+
},
|
| 722 |
+
{
|
| 723 |
+
"task": "openbookqa/mcq",
|
| 724 |
+
"n": 500,
|
| 725 |
+
"acc": 0.572,
|
| 726 |
+
"chance": 0.25,
|
| 727 |
+
"nll": 1.1920220527790952,
|
| 728 |
+
"brier": 0.6155284183368397,
|
| 729 |
+
"heldout": false,
|
| 730 |
+
"domain": "reasoning",
|
| 731 |
+
"ece": 0.21775564682483672,
|
| 732 |
+
"proof_v2 (card)": 0.292,
|
| 733 |
+
"\u0394 vs proof_v2": 0.27999999999999997
|
| 734 |
+
},
|
| 735 |
+
{
|
| 736 |
+
"task": "policy/access_control_transfer",
|
| 737 |
+
"n": 500,
|
| 738 |
+
"acc": 1.0,
|
| 739 |
+
"chance": 0.33333333333333326,
|
| 740 |
+
"nll": 0.002006467206090747,
|
| 741 |
+
"brier": 2.4728843418299015e-05,
|
| 742 |
+
"heldout": false,
|
| 743 |
+
"domain": "policy",
|
| 744 |
+
"ece": 0.001999327063560541,
|
| 745 |
+
"proof_v2 (card)": NaN,
|
| 746 |
+
"\u0394 vs proof_v2": NaN
|
| 747 |
+
},
|
| 748 |
+
{
|
| 749 |
+
"task": "policy/count_threshold_transfer",
|
| 750 |
+
"n": 500,
|
| 751 |
+
"acc": 0.88,
|
| 752 |
+
"chance": 0.10537052392052391,
|
| 753 |
+
"nll": 0.319376940273738,
|
| 754 |
+
"brier": 0.1809248790341789,
|
| 755 |
+
"heldout": false,
|
| 756 |
+
"domain": "policy",
|
| 757 |
+
"ece": 0.04758520478010175,
|
| 758 |
+
"proof_v2 (card)": NaN,
|
| 759 |
+
"\u0394 vs proof_v2": NaN
|
| 760 |
+
},
|
| 761 |
+
{
|
| 762 |
+
"task": "policy/free_shipping_transfer",
|
| 763 |
+
"n": 500,
|
| 764 |
+
"acc": 0.938,
|
| 765 |
+
"chance": 0.5,
|
| 766 |
+
"nll": 0.1287872482436942,
|
| 767 |
+
"brier": 0.08322479293591872,
|
| 768 |
+
"heldout": false,
|
| 769 |
+
"domain": "policy",
|
| 770 |
+
"ece": 0.041308605194091734,
|
| 771 |
+
"proof_v2 (card)": NaN,
|
| 772 |
+
"\u0394 vs proof_v2": NaN
|
| 773 |
+
},
|
| 774 |
+
{
|
| 775 |
+
"task": "policy/invoice_overdue_transfer",
|
| 776 |
+
"n": 500,
|
| 777 |
+
"acc": 0.824,
|
| 778 |
+
"chance": 0.5,
|
| 779 |
+
"nll": 0.3965394942490384,
|
| 780 |
+
"brier": 0.2504247577332596,
|
| 781 |
+
"heldout": false,
|
| 782 |
+
"domain": "policy",
|
| 783 |
+
"ece": 0.01985593855381013,
|
| 784 |
+
"proof_v2 (card)": NaN,
|
| 785 |
+
"\u0394 vs proof_v2": NaN
|
| 786 |
+
},
|
| 787 |
+
{
|
| 788 |
+
"task": "policy/invoice_total_transfer",
|
| 789 |
+
"n": 500,
|
| 790 |
+
"acc": 0.474,
|
| 791 |
+
"chance": 0.5,
|
| 792 |
+
"nll": 0.710681935429573,
|
| 793 |
+
"brier": 0.5164436840846652,
|
| 794 |
+
"heldout": false,
|
| 795 |
+
"domain": "policy",
|
| 796 |
+
"ece": 0.08196303224563599,
|
| 797 |
+
"proof_v2 (card)": NaN,
|
| 798 |
+
"\u0394 vs proof_v2": NaN
|
| 799 |
+
},
|
| 800 |
+
{
|
| 801 |
+
"task": "policy/refund_approval_transfer",
|
| 802 |
+
"n": 500,
|
| 803 |
+
"acc": 0.962,
|
| 804 |
+
"chance": 0.33333333333333326,
|
| 805 |
+
"nll": 0.10006511305032836,
|
| 806 |
+
"brier": 0.057969574647252664,
|
| 807 |
+
"heldout": false,
|
| 808 |
+
"domain": "policy",
|
| 809 |
+
"ece": 0.018834196925163266,
|
| 810 |
+
"proof_v2 (card)": NaN,
|
| 811 |
+
"\u0394 vs proof_v2": NaN
|
| 812 |
+
},
|
| 813 |
+
{
|
| 814 |
+
"task": "policy/return_window_transfer",
|
| 815 |
+
"n": 500,
|
| 816 |
+
"acc": 1.0,
|
| 817 |
+
"chance": 0.33333333333333326,
|
| 818 |
+
"nll": 0.02710788446944207,
|
| 819 |
+
"brier": 0.004185976644178241,
|
| 820 |
+
"heldout": false,
|
| 821 |
+
"domain": "policy",
|
| 822 |
+
"ece": 0.02567652928829193,
|
| 823 |
+
"proof_v2 (card)": NaN,
|
| 824 |
+
"\u0394 vs proof_v2": NaN
|
| 825 |
+
},
|
| 826 |
+
{
|
| 827 |
+
"task": "policy/sla_urgency_transfer",
|
| 828 |
+
"n": 500,
|
| 829 |
+
"acc": 0.606,
|
| 830 |
+
"chance": 0.25,
|
| 831 |
+
"nll": 0.8170870101451874,
|
| 832 |
+
"brier": 0.5123369210022903,
|
| 833 |
+
"heldout": false,
|
| 834 |
+
"domain": "policy",
|
| 835 |
+
"ece": 0.13068743020296097,
|
| 836 |
+
"proof_v2 (card)": NaN,
|
| 837 |
+
"\u0394 vs proof_v2": NaN
|
| 838 |
+
},
|
| 839 |
+
{
|
| 840 |
+
"task": "policy/table_compare_transfer",
|
| 841 |
+
"n": 500,
|
| 842 |
+
"acc": 0.928,
|
| 843 |
+
"chance": 0.5,
|
| 844 |
+
"nll": 0.2105378701629961,
|
| 845 |
+
"brier": 0.11898007306418568,
|
| 846 |
+
"heldout": false,
|
| 847 |
+
"domain": "policy",
|
| 848 |
+
"ece": 0.04929606854915622,
|
| 849 |
+
"proof_v2 (card)": NaN,
|
| 850 |
+
"\u0394 vs proof_v2": NaN
|
| 851 |
+
},
|
| 852 |
+
{
|
| 853 |
+
"task": "policy/table_count_transfer",
|
| 854 |
+
"n": 500,
|
| 855 |
+
"acc": 0.324,
|
| 856 |
+
"chance": 0.12114228549228548,
|
| 857 |
+
"nll": 3.24442003638536,
|
| 858 |
+
"brier": 1.1488026091975905,
|
| 859 |
+
"heldout": false,
|
| 860 |
+
"domain": "policy",
|
| 861 |
+
"ece": 0.5407210965156555,
|
| 862 |
+
"proof_v2 (card)": NaN,
|
| 863 |
+
"\u0394 vs proof_v2": NaN
|
| 864 |
+
},
|
| 865 |
+
{
|
| 866 |
+
"task": "policy/table_extreme_transfer",
|
| 867 |
+
"n": 500,
|
| 868 |
+
"acc": 0.932,
|
| 869 |
+
"chance": 0.12210815295815294,
|
| 870 |
+
"nll": 0.268062693382595,
|
| 871 |
+
"brier": 0.11322196827103552,
|
| 872 |
+
"heldout": false,
|
| 873 |
+
"domain": "policy",
|
| 874 |
+
"ece": 0.05259446609020234,
|
| 875 |
+
"proof_v2 (card)": NaN,
|
| 876 |
+
"\u0394 vs proof_v2": NaN
|
| 877 |
+
},
|
| 878 |
+
{
|
| 879 |
+
"task": "prompt_injections/detect",
|
| 880 |
+
"n": 116,
|
| 881 |
+
"acc": 0.5948275862068966,
|
| 882 |
+
"chance": 0.5,
|
| 883 |
+
"nll": 1.1838965486195179,
|
| 884 |
+
"brier": 0.6826645118148511,
|
| 885 |
+
"heldout": true,
|
| 886 |
+
"domain": "guardrails",
|
| 887 |
+
"ece": 0.33272201953263114,
|
| 888 |
+
"proof_v2 (card)": NaN,
|
| 889 |
+
"\u0394 vs proof_v2": NaN
|
| 890 |
+
},
|
| 891 |
+
{
|
| 892 |
+
"task": "qasc/mcq",
|
| 893 |
+
"n": 500,
|
| 894 |
+
"acc": 0.986,
|
| 895 |
+
"chance": 0.125,
|
| 896 |
+
"nll": 0.028506858559036514,
|
| 897 |
+
"brier": 0.017274368003708154,
|
| 898 |
+
"heldout": false,
|
| 899 |
+
"domain": "reasoning",
|
| 900 |
+
"ece": 0.007774531126022318,
|
| 901 |
+
"proof_v2 (card)": NaN,
|
| 902 |
+
"\u0394 vs proof_v2": NaN
|
| 903 |
+
},
|
| 904 |
+
{
|
| 905 |
+
"task": "sciq/mcq",
|
| 906 |
+
"n": 498,
|
| 907 |
+
"acc": 0.9538152610441767,
|
| 908 |
+
"chance": 0.25,
|
| 909 |
+
"nll": 0.1506263591345834,
|
| 910 |
+
"brier": 0.06985463059035983,
|
| 911 |
+
"heldout": false,
|
| 912 |
+
"domain": "reasoning",
|
| 913 |
+
"ece": 0.020590302994452258,
|
| 914 |
+
"proof_v2 (card)": 0.692,
|
| 915 |
+
"\u0394 vs proof_v2": 0.26181526104417674
|
| 916 |
+
},
|
| 917 |
+
{
|
| 918 |
+
"task": "scitail/support",
|
| 919 |
+
"n": 500,
|
| 920 |
+
"acc": 0.962,
|
| 921 |
+
"chance": 0.5,
|
| 922 |
+
"nll": 0.11612921302905306,
|
| 923 |
+
"brier": 0.05990850379879065,
|
| 924 |
+
"heldout": false,
|
| 925 |
+
"domain": "reasoning",
|
| 926 |
+
"ece": 0.03015338957309728,
|
| 927 |
+
"proof_v2 (card)": NaN,
|
| 928 |
+
"\u0394 vs proof_v2": NaN
|
| 929 |
+
},
|
| 930 |
+
{
|
| 931 |
+
"task": "snli/contradicts",
|
| 932 |
+
"n": 500,
|
| 933 |
+
"acc": 0.99,
|
| 934 |
+
"chance": 0.33333333333333326,
|
| 935 |
+
"nll": 0.042843449617153966,
|
| 936 |
+
"brier": 0.017027529190616314,
|
| 937 |
+
"heldout": false,
|
| 938 |
+
"domain": "reasoning",
|
| 939 |
+
"ece": 0.018511961221695038,
|
| 940 |
+
"proof_v2 (card)": 0.892,
|
| 941 |
+
"\u0394 vs proof_v2": 0.09799999999999998
|
| 942 |
+
},
|
| 943 |
+
{
|
| 944 |
+
"task": "snli/must_be_true",
|
| 945 |
+
"n": 500,
|
| 946 |
+
"acc": 0.986,
|
| 947 |
+
"chance": 0.33333333333333326,
|
| 948 |
+
"nll": 0.06753946821717545,
|
| 949 |
+
"brier": 0.02560326154020915,
|
| 950 |
+
"heldout": false,
|
| 951 |
+
"domain": "reasoning",
|
| 952 |
+
"ece": 0.043202445626258815,
|
| 953 |
+
"proof_v2 (card)": 0.908,
|
| 954 |
+
"\u0394 vs proof_v2": 0.07799999999999996
|
| 955 |
+
},
|
| 956 |
+
{
|
| 957 |
+
"task": "snli/nli",
|
| 958 |
+
"n": 988,
|
| 959 |
+
"acc": 0.8947368421052632,
|
| 960 |
+
"chance": 0.33333333333333326,
|
| 961 |
+
"nll": 0.3554420507991845,
|
| 962 |
+
"brier": 0.17963376213248355,
|
| 963 |
+
"heldout": false,
|
| 964 |
+
"domain": "reasoning",
|
| 965 |
+
"ece": 0.10791712172842222,
|
| 966 |
+
"proof_v2 (card)": NaN,
|
| 967 |
+
"\u0394 vs proof_v2": NaN
|
| 968 |
+
},
|
| 969 |
+
{
|
| 970 |
+
"task": "sst5/score",
|
| 971 |
+
"n": 500,
|
| 972 |
+
"acc": 0.406,
|
| 973 |
+
"chance": 0.2,
|
| 974 |
+
"nll": 1.3002369542717933,
|
| 975 |
+
"brier": 0.6888741553408958,
|
| 976 |
+
"heldout": true,
|
| 977 |
+
"domain": "classification",
|
| 978 |
+
"ece": 0.068851743131876,
|
| 979 |
+
"proof_v2 (card)": NaN,
|
| 980 |
+
"\u0394 vs proof_v2": NaN
|
| 981 |
+
},
|
| 982 |
+
{
|
| 983 |
+
"task": "tev1_test/ag_news",
|
| 984 |
+
"n": 150,
|
| 985 |
+
"acc": 0.92,
|
| 986 |
+
"chance": 0.25,
|
| 987 |
+
"nll": 0.2815313008365532,
|
| 988 |
+
"brier": 0.13456139206846243,
|
| 989 |
+
"heldout": true,
|
| 990 |
+
"domain": "tev1_benchmark",
|
| 991 |
+
"ece": 0.09770219445228576,
|
| 992 |
+
"proof_v2 (card)": NaN,
|
| 993 |
+
"\u0394 vs proof_v2": NaN
|
| 994 |
+
},
|
| 995 |
+
{
|
| 996 |
+
"task": "tev1_test/banking77",
|
| 997 |
+
"n": 200,
|
| 998 |
+
"acc": 0.72,
|
| 999 |
+
"chance": 0.17791666666666664,
|
| 1000 |
+
"nll": 0.777236735031438,
|
| 1001 |
+
"brier": 0.3902663054271286,
|
| 1002 |
+
"heldout": true,
|
| 1003 |
+
"domain": "tev1_benchmark",
|
| 1004 |
+
"ece": 0.13721376925706863,
|
| 1005 |
+
"proof_v2 (card)": NaN,
|
| 1006 |
+
"\u0394 vs proof_v2": NaN
|
| 1007 |
+
},
|
| 1008 |
+
{
|
| 1009 |
+
"task": "tev1_test/boolq",
|
| 1010 |
+
"n": 200,
|
| 1011 |
+
"acc": 0.835,
|
| 1012 |
+
"chance": 0.5,
|
| 1013 |
+
"nll": 0.40902234488283284,
|
| 1014 |
+
"brier": 0.24753055362669119,
|
| 1015 |
+
"heldout": true,
|
| 1016 |
+
"domain": "tev1_benchmark",
|
| 1017 |
+
"ece": 0.04955918818712238,
|
| 1018 |
+
"proof_v2 (card)": NaN,
|
| 1019 |
+
"\u0394 vs proof_v2": NaN
|
| 1020 |
+
},
|
| 1021 |
+
{
|
| 1022 |
+
"task": "tev1_test/mnli",
|
| 1023 |
+
"n": 300,
|
| 1024 |
+
"acc": 0.7233333333333334,
|
| 1025 |
+
"chance": 0.3333333333333333,
|
| 1026 |
+
"nll": 0.6736717029288412,
|
| 1027 |
+
"brier": 0.39782655881245704,
|
| 1028 |
+
"heldout": true,
|
| 1029 |
+
"domain": "tev1_benchmark",
|
| 1030 |
+
"ece": 0.07533363938331603,
|
| 1031 |
+
"proof_v2 (card)": NaN,
|
| 1032 |
+
"\u0394 vs proof_v2": NaN
|
| 1033 |
+
},
|
| 1034 |
+
{
|
| 1035 |
+
"task": "tev1_test/policy",
|
| 1036 |
+
"n": 1200,
|
| 1037 |
+
"acc": 0.5291666666666667,
|
| 1038 |
+
"chance": 0.3333333333333333,
|
| 1039 |
+
"nll": 0.8761954319352905,
|
| 1040 |
+
"brier": 0.5462664224307302,
|
| 1041 |
+
"heldout": true,
|
| 1042 |
+
"domain": "tev1_benchmark",
|
| 1043 |
+
"ece": 0.05868758827447889,
|
| 1044 |
+
"proof_v2 (card)": NaN,
|
| 1045 |
+
"\u0394 vs proof_v2": NaN
|
| 1046 |
+
},
|
| 1047 |
+
{
|
| 1048 |
+
"task": "tev1_test/routing",
|
| 1049 |
+
"n": 600,
|
| 1050 |
+
"acc": 0.4583333333333333,
|
| 1051 |
+
"chance": 0.2,
|
| 1052 |
+
"nll": 1.120521725914441,
|
| 1053 |
+
"brier": 0.5789269511485964,
|
| 1054 |
+
"heldout": true,
|
| 1055 |
+
"domain": "tev1_benchmark",
|
| 1056 |
+
"ece": 0.037736288358767814,
|
| 1057 |
+
"proof_v2 (card)": NaN,
|
| 1058 |
+
"\u0394 vs proof_v2": NaN
|
| 1059 |
+
},
|
| 1060 |
+
{
|
| 1061 |
+
"task": "tev1_test/sst5",
|
| 1062 |
+
"n": 150,
|
| 1063 |
+
"acc": 0.3933333333333333,
|
| 1064 |
+
"chance": 0.19999999999999996,
|
| 1065 |
+
"nll": 1.294863009850184,
|
| 1066 |
+
"brier": 0.6810328679494898,
|
| 1067 |
+
"heldout": true,
|
| 1068 |
+
"domain": "tev1_benchmark",
|
| 1069 |
+
"ece": 0.1081918982664744,
|
| 1070 |
+
"proof_v2 (card)": NaN,
|
| 1071 |
+
"\u0394 vs proof_v2": NaN
|
| 1072 |
+
},
|
| 1073 |
+
{
|
| 1074 |
+
"task": "triage/support_email",
|
| 1075 |
+
"n": 505,
|
| 1076 |
+
"acc": 0.9445544554455445,
|
| 1077 |
+
"chance": 0.32333333333333325,
|
| 1078 |
+
"nll": 0.15705669974757994,
|
| 1079 |
+
"brier": 0.07848840114983771,
|
| 1080 |
+
"heldout": false,
|
| 1081 |
+
"domain": "support",
|
| 1082 |
+
"ece": 0.045869263741049465,
|
| 1083 |
+
"proof_v2 (card)": NaN,
|
| 1084 |
+
"\u0394 vs proof_v2": NaN
|
| 1085 |
+
},
|
| 1086 |
+
{
|
| 1087 |
+
"task": "typed_decisions/agent_trace_observability",
|
| 1088 |
+
"n": 500,
|
| 1089 |
+
"acc": 0.734,
|
| 1090 |
+
"chance": 0.3,
|
| 1091 |
+
"nll": 0.8089503426551818,
|
| 1092 |
+
"brier": 0.45125534421065666,
|
| 1093 |
+
"heldout": false,
|
| 1094 |
+
"domain": "workflows",
|
| 1095 |
+
"ece": 0.22166652178764346,
|
| 1096 |
+
"proof_v2 (card)": NaN,
|
| 1097 |
+
"\u0394 vs proof_v2": NaN
|
| 1098 |
+
},
|
| 1099 |
+
{
|
| 1100 |
+
"task": "typed_decisions/customer_service",
|
| 1101 |
+
"n": 500,
|
| 1102 |
+
"acc": 0.764,
|
| 1103 |
+
"chance": 0.28,
|
| 1104 |
+
"nll": 0.7428050636351109,
|
| 1105 |
+
"brier": 0.40177602517263195,
|
| 1106 |
+
"heldout": false,
|
| 1107 |
+
"domain": "workflows",
|
| 1108 |
+
"ece": 0.20972245779633528,
|
| 1109 |
+
"proof_v2 (card)": NaN,
|
| 1110 |
+
"\u0394 vs proof_v2": NaN
|
| 1111 |
+
},
|
| 1112 |
+
{
|
| 1113 |
+
"task": "typed_decisions/invoice_processing",
|
| 1114 |
+
"n": 500,
|
| 1115 |
+
"acc": 0.826,
|
| 1116 |
+
"chance": 0.35,
|
| 1117 |
+
"nll": 0.5732684296518564,
|
| 1118 |
+
"brier": 0.2997603036136731,
|
| 1119 |
+
"heldout": false,
|
| 1120 |
+
"domain": "workflows",
|
| 1121 |
+
"ece": 0.19463082921504973,
|
| 1122 |
+
"proof_v2 (card)": NaN,
|
| 1123 |
+
"\u0394 vs proof_v2": NaN
|
| 1124 |
+
},
|
| 1125 |
+
{
|
| 1126 |
+
"task": "typed_decisions/security_incidents",
|
| 1127 |
+
"n": 500,
|
| 1128 |
+
"acc": 0.768,
|
| 1129 |
+
"chance": 0.34,
|
| 1130 |
+
"nll": 0.7502468670606613,
|
| 1131 |
+
"brier": 0.4238190137148829,
|
| 1132 |
+
"heldout": false,
|
| 1133 |
+
"domain": "workflows",
|
| 1134 |
+
"ece": 0.23508695220947262,
|
| 1135 |
+
"proof_v2 (card)": NaN,
|
| 1136 |
+
"\u0394 vs proof_v2": NaN
|
| 1137 |
+
},
|
| 1138 |
+
{
|
| 1139 |
+
"task": "winogrande/blank",
|
| 1140 |
+
"n": 500,
|
| 1141 |
+
"acc": 0.662,
|
| 1142 |
+
"chance": 0.5,
|
| 1143 |
+
"nll": 0.6651458515003323,
|
| 1144 |
+
"brier": 0.45323706188140617,
|
| 1145 |
+
"heldout": false,
|
| 1146 |
+
"domain": "reasoning",
|
| 1147 |
+
"ece": 0.1159557296037674,
|
| 1148 |
+
"proof_v2 (card)": NaN,
|
| 1149 |
+
"\u0394 vs proof_v2": NaN
|
| 1150 |
+
},
|
| 1151 |
+
{
|
| 1152 |
+
"task": "yelp/score",
|
| 1153 |
+
"n": 500,
|
| 1154 |
+
"acc": 0.668,
|
| 1155 |
+
"chance": 0.2,
|
| 1156 |
+
"nll": 0.7834262755662202,
|
| 1157 |
+
"brier": 0.4423740240858843,
|
| 1158 |
+
"heldout": false,
|
| 1159 |
+
"domain": "classification",
|
| 1160 |
+
"ece": 0.09068749487400055,
|
| 1161 |
+
"proof_v2 (card)": NaN,
|
| 1162 |
+
"\u0394 vs proof_v2": NaN
|
| 1163 |
+
}
|
| 1164 |
+
],
|
| 1165 |
+
"head_to_head": [
|
| 1166 |
+
{
|
| 1167 |
+
"task": "ag_news/topic",
|
| 1168 |
+
"n": 120,
|
| 1169 |
+
"LightDec": 0.85,
|
| 1170 |
+
"proof_v2": 0.45,
|
| 1171 |
+
"\u0394": 0.39999999999999997
|
| 1172 |
+
},
|
| 1173 |
+
{
|
| 1174 |
+
"task": "agentharm/refuse",
|
| 1175 |
+
"n": 120,
|
| 1176 |
+
"LightDec": 0.6666666666666666,
|
| 1177 |
+
"proof_v2": 0.45,
|
| 1178 |
+
"\u0394": 0.21666666666666662
|
| 1179 |
+
},
|
| 1180 |
+
{
|
| 1181 |
+
"task": "agenttrek/finish_now",
|
| 1182 |
+
"n": 120,
|
| 1183 |
+
"LightDec": 0.7833333333333333,
|
| 1184 |
+
"proof_v2": 0.25833333333333336,
|
| 1185 |
+
"\u0394": 0.5249999999999999
|
| 1186 |
+
},
|
| 1187 |
+
{
|
| 1188 |
+
"task": "agenttrek/next_action_type",
|
| 1189 |
+
"n": 120,
|
| 1190 |
+
"LightDec": 0.875,
|
| 1191 |
+
"proof_v2": 0.075,
|
| 1192 |
+
"\u0394": 0.8
|
| 1193 |
+
},
|
| 1194 |
+
{
|
| 1195 |
+
"task": "anli/nli",
|
| 1196 |
+
"n": 120,
|
| 1197 |
+
"LightDec": 0.5916666666666667,
|
| 1198 |
+
"proof_v2": 0.325,
|
| 1199 |
+
"\u0394": 0.26666666666666666
|
| 1200 |
+
},
|
| 1201 |
+
{
|
| 1202 |
+
"task": "aqua_rat/math",
|
| 1203 |
+
"n": 120,
|
| 1204 |
+
"LightDec": 0.35,
|
| 1205 |
+
"proof_v2": 0.24166666666666667,
|
| 1206 |
+
"\u0394": 0.10833333333333331
|
| 1207 |
+
},
|
| 1208 |
+
{
|
| 1209 |
+
"task": "arc_challenge/mcq",
|
| 1210 |
+
"n": 120,
|
| 1211 |
+
"LightDec": 0.5333333333333333,
|
| 1212 |
+
"proof_v2": 0.30833333333333335,
|
| 1213 |
+
"\u0394": 0.22499999999999998
|
| 1214 |
+
},
|
| 1215 |
+
{
|
| 1216 |
+
"task": "arc_easy/mcq",
|
| 1217 |
+
"n": 120,
|
| 1218 |
+
"LightDec": 0.65,
|
| 1219 |
+
"proof_v2": 0.4583333333333333,
|
| 1220 |
+
"\u0394": 0.1916666666666667
|
| 1221 |
+
},
|
| 1222 |
+
{
|
| 1223 |
+
"task": "banking77/intent",
|
| 1224 |
+
"n": 120,
|
| 1225 |
+
"LightDec": 0.925,
|
| 1226 |
+
"proof_v2": 0.8666666666666667,
|
| 1227 |
+
"\u0394": 0.05833333333333335
|
| 1228 |
+
},
|
| 1229 |
+
{
|
| 1230 |
+
"task": "bigclonebench/clone",
|
| 1231 |
+
"n": 120,
|
| 1232 |
+
"LightDec": 0.9666666666666667,
|
| 1233 |
+
"proof_v2": 0.24166666666666667,
|
| 1234 |
+
"\u0394": 0.725
|
| 1235 |
+
},
|
| 1236 |
+
{
|
| 1237 |
+
"task": "bitext/category",
|
| 1238 |
+
"n": 120,
|
| 1239 |
+
"LightDec": 1.0,
|
| 1240 |
+
"proof_v2": 0.8583333333333333,
|
| 1241 |
+
"\u0394": 0.14166666666666672
|
| 1242 |
+
},
|
| 1243 |
+
{
|
| 1244 |
+
"task": "bitext/route",
|
| 1245 |
+
"n": 120,
|
| 1246 |
+
"LightDec": 1.0,
|
| 1247 |
+
"proof_v2": 0.925,
|
| 1248 |
+
"\u0394": 0.07499999999999996
|
| 1249 |
+
},
|
| 1250 |
+
{
|
| 1251 |
+
"task": "boolq/yes_no",
|
| 1252 |
+
"n": 120,
|
| 1253 |
+
"LightDec": 0.825,
|
| 1254 |
+
"proof_v2": 0.6666666666666666,
|
| 1255 |
+
"\u0394": 0.15833333333333333
|
| 1256 |
+
},
|
| 1257 |
+
{
|
| 1258 |
+
"task": "civil_comments/toxic",
|
| 1259 |
+
"n": 120,
|
| 1260 |
+
"LightDec": 0.9333333333333333,
|
| 1261 |
+
"proof_v2": 0.35833333333333334,
|
| 1262 |
+
"\u0394": 0.575
|
| 1263 |
+
},
|
| 1264 |
+
{
|
| 1265 |
+
"task": "clinc150/intent",
|
| 1266 |
+
"n": 120,
|
| 1267 |
+
"LightDec": 0.975,
|
| 1268 |
+
"proof_v2": 0.6416666666666667,
|
| 1269 |
+
"\u0394": 0.33333333333333326
|
| 1270 |
+
},
|
| 1271 |
+
{
|
| 1272 |
+
"task": "codexglue/code_to_doc",
|
| 1273 |
+
"n": 120,
|
| 1274 |
+
"LightDec": 1.0,
|
| 1275 |
+
"proof_v2": 0.9333333333333333,
|
| 1276 |
+
"\u0394": 0.06666666666666665
|
| 1277 |
+
},
|
| 1278 |
+
{
|
| 1279 |
+
"task": "codexglue/doc_to_code",
|
| 1280 |
+
"n": 120,
|
| 1281 |
+
"LightDec": 0.9833333333333333,
|
| 1282 |
+
"proof_v2": 0.975,
|
| 1283 |
+
"\u0394": 0.008333333333333304
|
| 1284 |
+
},
|
| 1285 |
+
{
|
| 1286 |
+
"task": "codexglue/func_name",
|
| 1287 |
+
"n": 120,
|
| 1288 |
+
"LightDec": 0.9916666666666667,
|
| 1289 |
+
"proof_v2": 0.9166666666666666,
|
| 1290 |
+
"\u0394": 0.07500000000000007
|
| 1291 |
+
},
|
| 1292 |
+
{
|
| 1293 |
+
"task": "codexglue/lang_id",
|
| 1294 |
+
"n": 120,
|
| 1295 |
+
"LightDec": 1.0,
|
| 1296 |
+
"proof_v2": 1.0,
|
| 1297 |
+
"\u0394": 0.0
|
| 1298 |
+
},
|
| 1299 |
+
{
|
| 1300 |
+
"task": "commonsense_qa/mcq",
|
| 1301 |
+
"n": 120,
|
| 1302 |
+
"LightDec": 0.6833333333333333,
|
| 1303 |
+
"proof_v2": 0.4083333333333333,
|
| 1304 |
+
"\u0394": 0.275
|
| 1305 |
+
},
|
| 1306 |
+
{
|
| 1307 |
+
"task": "counsel/critique_quality",
|
| 1308 |
+
"n": 120,
|
| 1309 |
+
"LightDec": 0.6083333333333333,
|
| 1310 |
+
"proof_v2": 0.25833333333333336,
|
| 1311 |
+
"\u0394": 0.3499999999999999
|
| 1312 |
+
},
|
| 1313 |
+
{
|
| 1314 |
+
"task": "counsel/step_has_error",
|
| 1315 |
+
"n": 120,
|
| 1316 |
+
"LightDec": 0.825,
|
| 1317 |
+
"proof_v2": 0.75,
|
| 1318 |
+
"\u0394": 0.07499999999999996
|
| 1319 |
+
},
|
| 1320 |
+
{
|
| 1321 |
+
"task": "devign/vulnerability",
|
| 1322 |
+
"n": 120,
|
| 1323 |
+
"LightDec": 0.6666666666666666,
|
| 1324 |
+
"proof_v2": 0.5666666666666667,
|
| 1325 |
+
"\u0394": 0.09999999999999998
|
| 1326 |
+
},
|
| 1327 |
+
{
|
| 1328 |
+
"task": "emotion/6way",
|
| 1329 |
+
"n": 120,
|
| 1330 |
+
"LightDec": 0.49166666666666664,
|
| 1331 |
+
"proof_v2": 0.36666666666666664,
|
| 1332 |
+
"\u0394": 0.125
|
| 1333 |
+
},
|
| 1334 |
+
{
|
| 1335 |
+
"task": "gsm8k/math",
|
| 1336 |
+
"n": 120,
|
| 1337 |
+
"LightDec": 0.7333333333333333,
|
| 1338 |
+
"proof_v2": 0.125,
|
| 1339 |
+
"\u0394": 0.6083333333333333
|
| 1340 |
+
},
|
| 1341 |
+
{
|
| 1342 |
+
"task": "hellaswag/continuation",
|
| 1343 |
+
"n": 120,
|
| 1344 |
+
"LightDec": 0.6083333333333333,
|
| 1345 |
+
"proof_v2": 0.2916666666666667,
|
| 1346 |
+
"\u0394": 0.3166666666666666
|
| 1347 |
+
},
|
| 1348 |
+
{
|
| 1349 |
+
"task": "hotpotqa/comparison_yes_no",
|
| 1350 |
+
"n": 26,
|
| 1351 |
+
"LightDec": 0.9230769230769231,
|
| 1352 |
+
"proof_v2": 0.46153846153846156,
|
| 1353 |
+
"\u0394": 0.46153846153846156
|
| 1354 |
+
},
|
| 1355 |
+
{
|
| 1356 |
+
"task": "hotpotqa/retrieve",
|
| 1357 |
+
"n": 120,
|
| 1358 |
+
"LightDec": 0.875,
|
| 1359 |
+
"proof_v2": 0.2916666666666667,
|
| 1360 |
+
"\u0394": 0.5833333333333333
|
| 1361 |
+
},
|
| 1362 |
+
{
|
| 1363 |
+
"task": "humaneval/completion",
|
| 1364 |
+
"n": 119,
|
| 1365 |
+
"LightDec": 0.8487394957983193,
|
| 1366 |
+
"proof_v2": 0.5126050420168067,
|
| 1367 |
+
"\u0394": 0.33613445378151263
|
| 1368 |
+
},
|
| 1369 |
+
{
|
| 1370 |
+
"task": "jailbreak/detect",
|
| 1371 |
+
"n": 120,
|
| 1372 |
+
"LightDec": 0.9833333333333333,
|
| 1373 |
+
"proof_v2": 0.5666666666666667,
|
| 1374 |
+
"\u0394": 0.41666666666666663
|
| 1375 |
+
},
|
| 1376 |
+
{
|
| 1377 |
+
"task": "long/contract_clause",
|
| 1378 |
+
"n": 120,
|
| 1379 |
+
"LightDec": 1.0,
|
| 1380 |
+
"proof_v2": 0.23333333333333334,
|
| 1381 |
+
"\u0394": 0.7666666666666666
|
| 1382 |
+
},
|
| 1383 |
+
{
|
| 1384 |
+
"task": "long/email_thread",
|
| 1385 |
+
"n": 120,
|
| 1386 |
+
"LightDec": 1.0,
|
| 1387 |
+
"proof_v2": 0.05,
|
| 1388 |
+
"\u0394": 0.95
|
| 1389 |
+
},
|
| 1390 |
+
{
|
| 1391 |
+
"task": "long/service_log",
|
| 1392 |
+
"n": 120,
|
| 1393 |
+
"LightDec": 1.0,
|
| 1394 |
+
"proof_v2": 0.24166666666666667,
|
| 1395 |
+
"\u0394": 0.7583333333333333
|
| 1396 |
+
},
|
| 1397 |
+
{
|
| 1398 |
+
"task": "massive_en/intent",
|
| 1399 |
+
"n": 120,
|
| 1400 |
+
"LightDec": 0.95,
|
| 1401 |
+
"proof_v2": 0.7416666666666667,
|
| 1402 |
+
"\u0394": 0.20833333333333326
|
| 1403 |
+
},
|
| 1404 |
+
{
|
| 1405 |
+
"task": "mbpp/bugspot",
|
| 1406 |
+
"n": 120,
|
| 1407 |
+
"LightDec": 0.8583333333333333,
|
| 1408 |
+
"proof_v2": 0.575,
|
| 1409 |
+
"\u0394": 0.2833333333333333
|
| 1410 |
+
},
|
| 1411 |
+
{
|
| 1412 |
+
"task": "mbpp/solution",
|
| 1413 |
+
"n": 120,
|
| 1414 |
+
"LightDec": 1.0,
|
| 1415 |
+
"proof_v2": 0.85,
|
| 1416 |
+
"\u0394": 0.15000000000000002
|
| 1417 |
+
},
|
| 1418 |
+
{
|
| 1419 |
+
"task": "mmlu/mcq",
|
| 1420 |
+
"n": 120,
|
| 1421 |
+
"LightDec": 0.4083333333333333,
|
| 1422 |
+
"proof_v2": 0.30833333333333335,
|
| 1423 |
+
"\u0394": 0.09999999999999998
|
| 1424 |
+
},
|
| 1425 |
+
{
|
| 1426 |
+
"task": "mnli/claim",
|
| 1427 |
+
"n": 120,
|
| 1428 |
+
"LightDec": 0.875,
|
| 1429 |
+
"proof_v2": 0.31666666666666665,
|
| 1430 |
+
"\u0394": 0.5583333333333333
|
| 1431 |
+
},
|
| 1432 |
+
{
|
| 1433 |
+
"task": "openbookqa/mcq",
|
| 1434 |
+
"n": 120,
|
| 1435 |
+
"LightDec": 0.65,
|
| 1436 |
+
"proof_v2": 0.2833333333333333,
|
| 1437 |
+
"\u0394": 0.3666666666666667
|
| 1438 |
+
},
|
| 1439 |
+
{
|
| 1440 |
+
"task": "policy/access_control_transfer",
|
| 1441 |
+
"n": 120,
|
| 1442 |
+
"LightDec": 1.0,
|
| 1443 |
+
"proof_v2": 0.23333333333333334,
|
| 1444 |
+
"\u0394": 0.7666666666666666
|
| 1445 |
+
},
|
| 1446 |
+
{
|
| 1447 |
+
"task": "policy/count_threshold_transfer",
|
| 1448 |
+
"n": 120,
|
| 1449 |
+
"LightDec": 0.8666666666666667,
|
| 1450 |
+
"proof_v2": 0.10833333333333334,
|
| 1451 |
+
"\u0394": 0.7583333333333333
|
| 1452 |
+
},
|
| 1453 |
+
{
|
| 1454 |
+
"task": "policy/free_shipping_transfer",
|
| 1455 |
+
"n": 120,
|
| 1456 |
+
"LightDec": 0.925,
|
| 1457 |
+
"proof_v2": 0.7083333333333334,
|
| 1458 |
+
"\u0394": 0.21666666666666667
|
| 1459 |
+
},
|
| 1460 |
+
{
|
| 1461 |
+
"task": "policy/invoice_overdue_transfer",
|
| 1462 |
+
"n": 120,
|
| 1463 |
+
"LightDec": 0.8083333333333333,
|
| 1464 |
+
"proof_v2": 0.6583333333333333,
|
| 1465 |
+
"\u0394": 0.15000000000000002
|
| 1466 |
+
},
|
| 1467 |
+
{
|
| 1468 |
+
"task": "policy/invoice_total_transfer",
|
| 1469 |
+
"n": 120,
|
| 1470 |
+
"LightDec": 0.45,
|
| 1471 |
+
"proof_v2": 0.525,
|
| 1472 |
+
"\u0394": -0.07500000000000001
|
| 1473 |
+
},
|
| 1474 |
+
{
|
| 1475 |
+
"task": "policy/refund_approval_transfer",
|
| 1476 |
+
"n": 120,
|
| 1477 |
+
"LightDec": 0.9583333333333334,
|
| 1478 |
+
"proof_v2": 0.325,
|
| 1479 |
+
"\u0394": 0.6333333333333333
|
| 1480 |
+
},
|
| 1481 |
+
{
|
| 1482 |
+
"task": "policy/return_window_transfer",
|
| 1483 |
+
"n": 120,
|
| 1484 |
+
"LightDec": 1.0,
|
| 1485 |
+
"proof_v2": 0.3416666666666667,
|
| 1486 |
+
"\u0394": 0.6583333333333333
|
| 1487 |
+
},
|
| 1488 |
+
{
|
| 1489 |
+
"task": "policy/sla_urgency_transfer",
|
| 1490 |
+
"n": 120,
|
| 1491 |
+
"LightDec": 0.5166666666666667,
|
| 1492 |
+
"proof_v2": 0.44166666666666665,
|
| 1493 |
+
"\u0394": 0.07500000000000007
|
| 1494 |
+
},
|
| 1495 |
+
{
|
| 1496 |
+
"task": "policy/table_compare_transfer",
|
| 1497 |
+
"n": 120,
|
| 1498 |
+
"LightDec": 0.9166666666666666,
|
| 1499 |
+
"proof_v2": 0.49166666666666664,
|
| 1500 |
+
"\u0394": 0.425
|
| 1501 |
+
},
|
| 1502 |
+
{
|
| 1503 |
+
"task": "policy/table_count_transfer",
|
| 1504 |
+
"n": 120,
|
| 1505 |
+
"LightDec": 0.31666666666666665,
|
| 1506 |
+
"proof_v2": 0.15833333333333333,
|
| 1507 |
+
"\u0394": 0.15833333333333333
|
| 1508 |
+
},
|
| 1509 |
+
{
|
| 1510 |
+
"task": "policy/table_extreme_transfer",
|
| 1511 |
+
"n": 120,
|
| 1512 |
+
"LightDec": 0.925,
|
| 1513 |
+
"proof_v2": 0.14166666666666666,
|
| 1514 |
+
"\u0394": 0.7833333333333334
|
| 1515 |
+
},
|
| 1516 |
+
{
|
| 1517 |
+
"task": "prompt_injections/detect",
|
| 1518 |
+
"n": 116,
|
| 1519 |
+
"LightDec": 0.5948275862068966,
|
| 1520 |
+
"proof_v2": 0.5431034482758621,
|
| 1521 |
+
"\u0394": 0.051724137931034475
|
| 1522 |
+
},
|
| 1523 |
+
{
|
| 1524 |
+
"task": "qasc/mcq",
|
| 1525 |
+
"n": 120,
|
| 1526 |
+
"LightDec": 0.9916666666666667,
|
| 1527 |
+
"proof_v2": 0.8583333333333333,
|
| 1528 |
+
"\u0394": 0.13333333333333341
|
| 1529 |
+
},
|
| 1530 |
+
{
|
| 1531 |
+
"task": "sciq/mcq",
|
| 1532 |
+
"n": 120,
|
| 1533 |
+
"LightDec": 0.9666666666666667,
|
| 1534 |
+
"proof_v2": 0.8666666666666667,
|
| 1535 |
+
"\u0394": 0.09999999999999998
|
| 1536 |
+
},
|
| 1537 |
+
{
|
| 1538 |
+
"task": "scitail/support",
|
| 1539 |
+
"n": 120,
|
| 1540 |
+
"LightDec": 0.9583333333333334,
|
| 1541 |
+
"proof_v2": 0.425,
|
| 1542 |
+
"\u0394": 0.5333333333333334
|
| 1543 |
+
},
|
| 1544 |
+
{
|
| 1545 |
+
"task": "snli/contradicts",
|
| 1546 |
+
"n": 120,
|
| 1547 |
+
"LightDec": 1.0,
|
| 1548 |
+
"proof_v2": 0.35833333333333334,
|
| 1549 |
+
"\u0394": 0.6416666666666666
|
| 1550 |
+
},
|
| 1551 |
+
{
|
| 1552 |
+
"task": "snli/must_be_true",
|
| 1553 |
+
"n": 120,
|
| 1554 |
+
"LightDec": 0.9916666666666667,
|
| 1555 |
+
"proof_v2": 0.925,
|
| 1556 |
+
"\u0394": 0.06666666666666665
|
| 1557 |
+
},
|
| 1558 |
+
{
|
| 1559 |
+
"task": "snli/nli",
|
| 1560 |
+
"n": 120,
|
| 1561 |
+
"LightDec": 0.9083333333333333,
|
| 1562 |
+
"proof_v2": 0.3416666666666667,
|
| 1563 |
+
"\u0394": 0.5666666666666667
|
| 1564 |
+
},
|
| 1565 |
+
{
|
| 1566 |
+
"task": "sst5/score",
|
| 1567 |
+
"n": 120,
|
| 1568 |
+
"LightDec": 0.38333333333333336,
|
| 1569 |
+
"proof_v2": 0.2833333333333333,
|
| 1570 |
+
"\u0394": 0.10000000000000003
|
| 1571 |
+
},
|
| 1572 |
+
{
|
| 1573 |
+
"task": "tev1_test/ag_news",
|
| 1574 |
+
"n": 120,
|
| 1575 |
+
"LightDec": 0.925,
|
| 1576 |
+
"proof_v2": 0.425,
|
| 1577 |
+
"\u0394": 0.5
|
| 1578 |
+
},
|
| 1579 |
+
{
|
| 1580 |
+
"task": "tev1_test/banking77",
|
| 1581 |
+
"n": 120,
|
| 1582 |
+
"LightDec": 0.7166666666666667,
|
| 1583 |
+
"proof_v2": 0.55,
|
| 1584 |
+
"\u0394": 0.16666666666666663
|
| 1585 |
+
},
|
| 1586 |
+
{
|
| 1587 |
+
"task": "tev1_test/boolq",
|
| 1588 |
+
"n": 120,
|
| 1589 |
+
"LightDec": 0.8416666666666667,
|
| 1590 |
+
"proof_v2": 0.5583333333333333,
|
| 1591 |
+
"\u0394": 0.2833333333333333
|
| 1592 |
+
},
|
| 1593 |
+
{
|
| 1594 |
+
"task": "tev1_test/mnli",
|
| 1595 |
+
"n": 120,
|
| 1596 |
+
"LightDec": 0.75,
|
| 1597 |
+
"proof_v2": 0.39166666666666666,
|
| 1598 |
+
"\u0394": 0.35833333333333334
|
| 1599 |
+
},
|
| 1600 |
+
{
|
| 1601 |
+
"task": "tev1_test/policy",
|
| 1602 |
+
"n": 120,
|
| 1603 |
+
"LightDec": 0.5,
|
| 1604 |
+
"proof_v2": 0.4083333333333333,
|
| 1605 |
+
"\u0394": 0.09166666666666667
|
| 1606 |
+
},
|
| 1607 |
+
{
|
| 1608 |
+
"task": "tev1_test/routing",
|
| 1609 |
+
"n": 120,
|
| 1610 |
+
"LightDec": 0.44166666666666665,
|
| 1611 |
+
"proof_v2": 0.2916666666666667,
|
| 1612 |
+
"\u0394": 0.14999999999999997
|
| 1613 |
+
},
|
| 1614 |
+
{
|
| 1615 |
+
"task": "tev1_test/sst5",
|
| 1616 |
+
"n": 120,
|
| 1617 |
+
"LightDec": 0.39166666666666666,
|
| 1618 |
+
"proof_v2": 0.275,
|
| 1619 |
+
"\u0394": 0.11666666666666664
|
| 1620 |
+
},
|
| 1621 |
+
{
|
| 1622 |
+
"task": "triage/support_email",
|
| 1623 |
+
"n": 120,
|
| 1624 |
+
"LightDec": 0.9583333333333334,
|
| 1625 |
+
"proof_v2": 0.39166666666666666,
|
| 1626 |
+
"\u0394": 0.5666666666666667
|
| 1627 |
+
},
|
| 1628 |
+
{
|
| 1629 |
+
"task": "typed_decisions/agent_trace_observability",
|
| 1630 |
+
"n": 120,
|
| 1631 |
+
"LightDec": 0.8083333333333333,
|
| 1632 |
+
"proof_v2": 0.3333333333333333,
|
| 1633 |
+
"\u0394": 0.47500000000000003
|
| 1634 |
+
},
|
| 1635 |
+
{
|
| 1636 |
+
"task": "typed_decisions/customer_service",
|
| 1637 |
+
"n": 120,
|
| 1638 |
+
"LightDec": 0.7416666666666667,
|
| 1639 |
+
"proof_v2": 0.2833333333333333,
|
| 1640 |
+
"\u0394": 0.45833333333333337
|
| 1641 |
+
},
|
| 1642 |
+
{
|
| 1643 |
+
"task": "typed_decisions/invoice_processing",
|
| 1644 |
+
"n": 120,
|
| 1645 |
+
"LightDec": 0.8166666666666667,
|
| 1646 |
+
"proof_v2": 0.475,
|
| 1647 |
+
"\u0394": 0.3416666666666667
|
| 1648 |
+
},
|
| 1649 |
+
{
|
| 1650 |
+
"task": "typed_decisions/security_incidents",
|
| 1651 |
+
"n": 120,
|
| 1652 |
+
"LightDec": 0.7583333333333333,
|
| 1653 |
+
"proof_v2": 0.5,
|
| 1654 |
+
"\u0394": 0.2583333333333333
|
| 1655 |
+
},
|
| 1656 |
+
{
|
| 1657 |
+
"task": "winogrande/blank",
|
| 1658 |
+
"n": 120,
|
| 1659 |
+
"LightDec": 0.6916666666666667,
|
| 1660 |
+
"proof_v2": 0.5583333333333333,
|
| 1661 |
+
"\u0394": 0.1333333333333333
|
| 1662 |
+
},
|
| 1663 |
+
{
|
| 1664 |
+
"task": "yelp/score",
|
| 1665 |
+
"n": 120,
|
| 1666 |
+
"LightDec": 0.625,
|
| 1667 |
+
"proof_v2": 0.325,
|
| 1668 |
+
"\u0394": 0.3
|
| 1669 |
+
}
|
| 1670 |
+
],
|
| 1671 |
+
"head_to_head_error": null,
|
| 1672 |
+
"v3_benchmarks": {
|
| 1673 |
+
"tev1": {
|
| 1674 |
+
"per_task": [
|
| 1675 |
+
{
|
| 1676 |
+
"task": "tev1_test/ag_news",
|
| 1677 |
+
"n": 150,
|
| 1678 |
+
"acc": 0.92,
|
| 1679 |
+
"overlaps_LightDec_training_source": true
|
| 1680 |
+
},
|
| 1681 |
+
{
|
| 1682 |
+
"task": "tev1_test/banking77",
|
| 1683 |
+
"n": 200,
|
| 1684 |
+
"acc": 0.72,
|
| 1685 |
+
"overlaps_LightDec_training_source": false
|
| 1686 |
+
},
|
| 1687 |
+
{
|
| 1688 |
+
"task": "tev1_test/boolq",
|
| 1689 |
+
"n": 200,
|
| 1690 |
+
"acc": 0.835,
|
| 1691 |
+
"overlaps_LightDec_training_source": true
|
| 1692 |
+
},
|
| 1693 |
+
{
|
| 1694 |
+
"task": "tev1_test/mnli",
|
| 1695 |
+
"n": 300,
|
| 1696 |
+
"acc": 0.7233333333333334,
|
| 1697 |
+
"overlaps_LightDec_training_source": true
|
| 1698 |
+
},
|
| 1699 |
+
{
|
| 1700 |
+
"task": "tev1_test/policy",
|
| 1701 |
+
"n": 1200,
|
| 1702 |
+
"acc": 0.5291666666666667,
|
| 1703 |
+
"overlaps_LightDec_training_source": false
|
| 1704 |
+
},
|
| 1705 |
+
{
|
| 1706 |
+
"task": "tev1_test/routing",
|
| 1707 |
+
"n": 600,
|
| 1708 |
+
"acc": 0.4583333333333333,
|
| 1709 |
+
"overlaps_LightDec_training_source": false
|
| 1710 |
+
},
|
| 1711 |
+
{
|
| 1712 |
+
"task": "tev1_test/sst5",
|
| 1713 |
+
"n": 150,
|
| 1714 |
+
"acc": 0.3933333333333333,
|
| 1715 |
+
"overlaps_LightDec_training_source": false
|
| 1716 |
+
}
|
| 1717 |
+
],
|
| 1718 |
+
"all": 0.5839285714285715,
|
| 1719 |
+
"no_overlap": 0.5176744186046511,
|
| 1720 |
+
"n": 2800,
|
| 1721 |
+
"published_tev1_4b": {
|
| 1722 |
+
"main decisions (1,000)": 0.88,
|
| 1723 |
+
"policy transfer (300)": 1.0
|
| 1724 |
+
}
|
| 1725 |
+
},
|
| 1726 |
+
"triage": {
|
| 1727 |
+
"emails": 101,
|
| 1728 |
+
"per_question": {
|
| 1729 |
+
"topic": 0.9900990099009901,
|
| 1730 |
+
"refund": 1.0,
|
| 1731 |
+
"breakage": 0.9900990099009901,
|
| 1732 |
+
"anger": 0.9207920792079208,
|
| 1733 |
+
"judgment": 0.8217821782178217
|
| 1734 |
+
},
|
| 1735 |
+
"pile_accuracy": 0.9405940594059405
|
| 1736 |
+
},
|
| 1737 |
+
"long_context": {
|
| 1738 |
+
"max_len": 2048,
|
| 1739 |
+
"per_task": [
|
| 1740 |
+
{
|
| 1741 |
+
"task": "long/contract_clause",
|
| 1742 |
+
"n": 150,
|
| 1743 |
+
"acc": 1.0,
|
| 1744 |
+
"chance": 0.19999999999999996
|
| 1745 |
+
},
|
| 1746 |
+
{
|
| 1747 |
+
"task": "long/email_thread",
|
| 1748 |
+
"n": 150,
|
| 1749 |
+
"acc": 1.0,
|
| 1750 |
+
"chance": 0.25
|
| 1751 |
+
},
|
| 1752 |
+
{
|
| 1753 |
+
"task": "long/service_log",
|
| 1754 |
+
"n": 150,
|
| 1755 |
+
"acc": 1.0,
|
| 1756 |
+
"chance": 0.19999999999999996
|
| 1757 |
+
}
|
| 1758 |
+
]
|
| 1759 |
+
}
|
| 1760 |
+
},
|
| 1761 |
+
"architecture": "FalconDec",
|
| 1762 |
+
"latency": {
|
| 1763 |
+
"gpu": "NVIDIA RTX PRO 6000 Blackwell Server Edition",
|
| 1764 |
+
"calls": [
|
| 1765 |
+
{
|
| 1766 |
+
"device": "cuda",
|
| 1767 |
+
"weights": "fp16",
|
| 1768 |
+
"questions": 1,
|
| 1769 |
+
"p50_ms": 10.05,
|
| 1770 |
+
"p95_ms": 10.43
|
| 1771 |
+
},
|
| 1772 |
+
{
|
| 1773 |
+
"device": "cuda",
|
| 1774 |
+
"weights": "fp16",
|
| 1775 |
+
"questions": 5,
|
| 1776 |
+
"p50_ms": 11.32,
|
| 1777 |
+
"p95_ms": 11.41
|
| 1778 |
+
},
|
| 1779 |
+
{
|
| 1780 |
+
"device": "cuda",
|
| 1781 |
+
"weights": "fp16",
|
| 1782 |
+
"questions": 10,
|
| 1783 |
+
"p50_ms": 12.57,
|
| 1784 |
+
"p95_ms": 12.69
|
| 1785 |
+
},
|
| 1786 |
+
{
|
| 1787 |
+
"device": "cpu fp32",
|
| 1788 |
+
"weights": "int8 file",
|
| 1789 |
+
"questions": 1,
|
| 1790 |
+
"p50_ms": 48.4,
|
| 1791 |
+
"p95_ms": 48.8
|
| 1792 |
+
}
|
| 1793 |
+
],
|
| 1794 |
+
"throughput_per_s": 2588.8502560238694
|
| 1795 |
+
},
|
| 1796 |
+
"env": {
|
| 1797 |
+
"torch": "2.9.0+cu130",
|
| 1798 |
+
"transformers": "5.17.0",
|
| 1799 |
+
"python": "3.12.12",
|
| 1800 |
+
"gpu": "NVIDIA RTX PRO 6000 Blackwell Server Edition"
|
| 1801 |
+
}
|
| 1802 |
+
}
|
model.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:c675a38babe2bb34a81efc13936a134b7596cc5b12da859bf28966b537659e0b
|
| 3 |
+
size 319325890
|
tokenizer/tokenizer.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
tokenizer/tokenizer_config.json
ADDED
|
@@ -0,0 +1,17 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"backend": "tokenizers",
|
| 3 |
+
"clean_up_tokenization_spaces": true,
|
| 4 |
+
"cls_token": "[CLS]",
|
| 5 |
+
"is_local": false,
|
| 6 |
+
"local_files_only": false,
|
| 7 |
+
"mask_token": "[MASK]",
|
| 8 |
+
"model_input_names": [
|
| 9 |
+
"input_ids",
|
| 10 |
+
"attention_mask"
|
| 11 |
+
],
|
| 12 |
+
"model_max_length": 8192,
|
| 13 |
+
"pad_token": "[PAD]",
|
| 14 |
+
"sep_token": "[SEP]",
|
| 15 |
+
"tokenizer_class": "TokenizersBackend",
|
| 16 |
+
"unk_token": "[UNK]"
|
| 17 |
+
}
|