Release Qev-9B v0.2.0: Principle and boundary training, MMLU-Pro 57.40

NOTICE

@@ -13,8 +13,9 @@ Adaptations include Qwen delimiters, text rendering, LoRA target selection, and
 inference cache forking. Modified files carry Qev adaptation notices. The original
 Kev license and copyright notice are retained in licenses/Kev-Apache-2.0.txt.
 
-Qev uses the Qwen3.5 model and tokenizer:
+Qev uses Qwen3.5 models and tokenizers:
   https://huggingface.co/Qwen/Qwen3.5-9B-Base
+  https://huggingface.co/Qwen/Qwen3.5-2B-Base
   Copyright 2026 Alibaba Cloud
   Apache License, Version 2.0
 
@@ -28,3 +29,8 @@ third-party notices, and per-question provenance with the prepared data.
 
 See THIRD_PARTY_NOTICES.md for licensing scope and external dependencies.
 Third-party datasets and services retain their own terms.
+
+Crafter visuals in the project gameplay recording:
+  https://github.com/danijar/crafter
+  Copyright (c) 2021 Danijar Hafner
+  MIT License; see licenses/Crafter-MIT.txt.

README.md

@@ -2,25 +2,25 @@
 license: apache-2.0
 base_model: Qwen/Qwen3.5-9B-Base
 base_model_relation: adapter
+datasets:
+  - AustinFu/Qev-train
 language:
-- en
-- zh
+  - en
+  - zh
 tags:
-- qev
-- qwen3_5
-- lora
-- decision-model
-- choice
-- noul
-- score
+  - qev
+  - qwen3_5
+  - lora
+  - decision-model
+  - choice
+  - noul
+  - score
 metrics:
-- accuracy
-datasets:
-- AustinFu/Qev-train
+  - accuracy
 ---
 
 <p align="center">
-  <img src="https://raw.githubusercontent.com/QiqianFu/Qev/main/assets/banner.svg" alt="Qev — decisions, grounded in Qwen" width="100%">
+  <img src="assets/banner.svg" alt="Qev — decisions, grounded in Qwen" width="100%">
 </p>
 
 # Qev-9B
@@ -29,7 +29,7 @@ datasets:
 
 [Source and documentation](https://github.com/QiqianFu/Qev) · [中文说明](https://github.com/QiqianFu/Qev/blob/main/README.zh-CN.md) · [Model weights](https://huggingface.co/AustinFu/Qev-9B)
 
-Qev-9B combines a general decision-training dataset with additional alignment and rule-compliance examples, repeated three times during the second half of training. It uses the full candidate-interaction architecture shown below.
+Qev-9B v0.2.0 combines general decision training with HelpSteer3 Principle judgments and 600 synthetic boundary questions in the main set. Additional alignment and document-rule examples are repeated three times during the second half of training. It uses the full candidate-interaction architecture shown below.
 
 ## Quick start
 
@@ -75,7 +75,7 @@ python -m qev.predict \
 ## Architecture and training
 
 <p align="center">
-  <img src="https://raw.githubusercontent.com/QiqianFu/Qev/main/assets/architecture.svg" alt="Qev encodes context, questions and answer options, then scores the options with its decision head." width="100%">
+  <img src="assets/architecture.svg" alt="Qev encodes context, questions and answer options, then scores the options with its decision head." width="100%">
 </p>
 
 | Field | Released model |
@@ -86,33 +86,34 @@ python -m qev.predict \
 | Backbone interaction | `last-full-attention` |
 | Computation | BF16 backbone; FP32 decision head and key reductions |
 | Stored adaptation tensors | FP32 |
-| Main training partition | 34,546 records |
+| Model release | v0.2.0 |
+| Main training partition | 39,605 records, including 4,459 Principle judgments and 600 synthetic boundary questions |
 | Late partition | 1,419 alignment records + 364 rule-compliance judgments |
 | Training schedule | Two epochs, global batch 32, seed 17 |
 | Late mixing | Starts halfway through main training; late examples repeat three times |
-| Selected checkpoint | Step 2327 |
+| Selected checkpoint | Step 2643 |
 
-The main and late partitions intentionally share 249 replay records. [Qev-train](https://huggingface.co/datasets/AustinFu/Qev-train) publishes 1,842 synthetic alignment, rule-compliance and world-knowledge examples with original hard labels and generation documentation. The complete mixed training corpus is not distributed. [Training guide](https://github.com/QiqianFu/Qev/blob/main/docs/training.md) · [Data recipe](https://github.com/QiqianFu/Qev/blob/main/docs/data.md).
+The main and late partitions intentionally share 249 replay records. [Qev-train](https://huggingface.co/datasets/AustinFu/Qev-train) publishes 2,442 synthetic alignment, rule-compliance, world-knowledge and HelpSteer3-derived boundary examples, with generation methods and source-specific licenses. The complete mixed training corpus is not distributed. [Training guide](https://github.com/QiqianFu/Qev/blob/main/docs/training.md) · [Data recipe](https://github.com/QiqianFu/Qev/blob/main/docs/data.md).
 
 ## Evaluation
 
 **Qev-9B uses BF16 backbone computation; Kev-9B uses FP32.**
 
 <p align="center">
-  <img src="https://raw.githubusercontent.com/QiqianFu/Qev/main/assets/evaluation.svg" alt="Qev-9B and Qev-2B accuracy compared with their Qwen3.5 base models on seven benchmarks." width="100%">
+  <img src="assets/evaluation.svg" alt="Qev-9B and Qev-2B accuracy compared with their Qwen3.5 base models on seven benchmarks." width="100%">
 </p>
 
 | Benchmark | Jev (reference) | Qwen3.5-9B-Base | Qev-9B | Kev-9B |
 |---|---:|---:|---:|---:|
 | Decision development · clean | 84.49 | 77.69 | **87.42** | 87.18 |
 | Transfer development · clean | 85.67 | 74.39 | **83.99** | 82.16 |
-| MMLU-Pro · 1,000 | 83.50 | 50.40 | **54.60** | 51.10 |
-| SemIf · 144 handwritten | 96.53 | 90.28 | **93.75** | 90.97 |
-| scienthoon · 873 | 75.26 | 68.84 | 72.28 | **75.49** |
-| WANLI · 256 | 75.78 | 67.97 | **72.66** | 70.31 |
-| JevBench public · 231 | 85.71 | 75.76 | **81.39** | 75.76 |
+| MMLU-Pro · 1,000 | 83.50 | 50.40 | **57.40** | 51.10 |
+| SemIf · 144 handwritten | 96.53 | 90.28 | **93.06** | 90.97 |
+| scienthoon · 873 | 75.26 | 68.84 | 71.02 | **75.49** |
+| WANLI · 256 | 75.78 | 67.97 | **71.09** | 70.31 |
+| JevBench public · 231 | 85.71 | 75.76 | **83.12** | 75.76 |
 
-Accuracy (%). Bold compares Qev with Kev. JevBench is public-set accuracy: Qev answers 188 of 231 questions correctly. It is not the official JevBench composite score.
+Accuracy (%). Bold compares Qev with Kev.
 
 These are the recorded results for the released checkpoint, using full causal reference execution. The selected model is a single seed, and public benchmarks were observed during research iteration. The comparisons do not isolate architecture gains. [Results, sources, and reproduction commands](https://github.com/QiqianFu/Qev/blob/main/docs/evaluation.md).
 
@@ -132,4 +133,4 @@ The adaptation tensors are byte-identical to the selected research checkpoint. Q
 
 Qev supports research and development of routing, rule judgments, and rubric ratings over explicit options. Evaluate it on your application's inputs and decision thresholds. Probabilities depend on the supplied options; calibration and production reliability have not been established.
 
-Qev's code, adaptation weights, documentation and original illustrations use Apache-2.0. Qwen models and the bundled tokenizer retain Alibaba Cloud's Apache-2.0 license. Qev adapts conventions from [Jared Palmer's Kev](https://github.com/jaredpalmer/kev), whose Apache-2.0 license and attribution are retained. JevBench tasks and external dependencies retain their own terms; the training corpus is not redistributed. See the included `LICENSE`, `NOTICE`, `THIRD_PARTY_NOTICES.md`, and `licenses/` for the full texts and scope.
+Qev's code, adaptation weights, documentation and original illustrations use Apache-2.0. Qwen models and the bundled tokenizer retain Alibaba Cloud's Apache-2.0 license. Qev adapts conventions from [Jared Palmer's Kev](https://github.com/jaredpalmer/kev), whose Apache-2.0 license and attribution are retained. JevBench tasks and external dependencies retain their own terms; the complete mixed training corpus is not bundled. The synthetic subset is published separately as Qev-train under its component-specific licenses. See the included `LICENSE`, `NOTICE`, `THIRD_PARTY_NOTICES.md`, and `licenses/` for the full texts and scope.

THIRD_PARTY_NOTICES.md

@@ -13,12 +13,13 @@ Third-party materials retain the terms listed below. The Qev license does not re
 | Material | Source and copyright | License and scope |
 |---|---|---|
 | Kev conventions and adapted implementation | [Jared Palmer's Kev](https://github.com/jaredpalmer/kev), Copyright 2026 Jared Palmer | [Apache-2.0, original notice retained](licenses/Kev-Apache-2.0.txt). Qev adapts delimiter and rendering conventions, LoRA targets, and cache forking. Modified source files identify the Qev adaptation. |
-| Qwen model and tokenizer | [Qwen3.5-9B-Base](https://huggingface.co/Qwen/Qwen3.5-9B-Base), Copyright 2026 Alibaba Cloud | [Apache-2.0](licenses/Qwen3.5-Apache-2.0.txt). The tokenizer is included in the Hugging Face checkpoint; base weights are downloaded separately. |
+| Qwen models and tokenizers | [Qwen3.5-9B-Base](https://huggingface.co/Qwen/Qwen3.5-9B-Base) and [Qwen3.5-2B-Base](https://huggingface.co/Qwen/Qwen3.5-2B-Base), Copyright 2026 Alibaba Cloud | [Apache-2.0](licenses/Qwen3.5-Apache-2.0.txt). Tokenizers accompany the corresponding Qev checkpoints; base weights are downloaded separately. |
 | JevBench public evaluation tasks | [JevBench](https://github.com/fstandhartinger/jevbench), Copyright 2026 Florian Standhartinger and contributors | [MIT](licenses/JevBench-MIT.txt), with [upstream third-party notices](licenses/JevBench-THIRD-PARTY.md). Tasks are downloaded by the optional preparation script. The original license, third-party notice, and per-question provenance accompany the prepared dataset. |
+| Crafter visuals in the gameplay recording | [Crafter](https://github.com/danijar/crafter), Copyright 2021 Danijar Hafner | [MIT](licenses/Crafter-MIT.txt). The gameplay recording contains the environment's visual assets; the recording and Qev decision overlay were produced for this project. |
 
 The reviewed Kev version is [557598f](https://github.com/jaredpalmer/kev/tree/557598fced1dada75dfbf36ed144dce309ac6ceb); the evaluation data use [JevBench v1.4.2](https://github.com/fstandhartinger/jevbench/tree/1df665e3956d7aab7fa0208ff6c4f2d8557f9f90). The 231 downloaded public tasks carry MIT in their individual provenance records; upstream notices describe the scope of other JevBench materials and evaluated services.
 
-The README layout takes inspiration from [JevAny](https://github.com/weitianxin/JevAny). Qev's SVG illustrations and chart code are original; this release includes no JevAny game code or recordings.
+The README layout takes inspiration from [JevAny](https://github.com/weitianxin/JevAny). Qev's SVG illustrations and chart code are original. The GIF and MP4 demos are this project's recordings of 9B research models. The Crafter research integration used JevAny's environment adapters; this source package does not vendor those adapters or JevAny's bundled recordings.
 
 ## Dependencies installed separately
 
@@ -32,6 +33,10 @@ Their source code and dependency distributions are not vendored in Qev's Python
 
 ## Training data
 
-The complete research training mixture is not distributed here. Public weight availability does not make the training data available under Apache-2.0. Any future dataset release needs its own source-specific permissions and attribution. The Wikipedia-derived training subset, for example, carries CC BY-SA attribution where applicable.
+The [Qev-train dataset](https://huggingface.co/datasets/AustinFu/Qev-train) publishes 2,442 synthetic training examples separately from this code repository. Its 1,534 original alignment and rule-compliance examples use Apache-2.0 where copyright applies. Its 308 Wikipedia-grounded knowledge examples retain CC BY-SA 4.0, with article versions, contributor links and changes recorded in `ATTRIBUTION.jsonl`. Its 600 boundary examples adapt NVIDIA HelpSteer3 Preference contexts and retain [CC BY 4.0](licenses/HelpSteer3-CC-BY-4.0.txt), with source lineage and changes in `HELPSTEER3_ATTRIBUTION.jsonl`. The [dataset license](https://huggingface.co/datasets/AustinFu/Qev-train/blob/main/LICENSE.md) identifies the scope of each component; combining them does not relicense the derived components under Apache-2.0.
+
+Qev-9B v0.2.0 also uses 4,459 Principle judgments from [NVIDIA HelpSteer3](https://huggingface.co/datasets/nvidia/HelpSteer3), by Zhilin Wang and collaborators, under CC BY 4.0. The source revision is `f6d145777bcbde96137596340fab89793acd1031`; those original judgments are not part of the synthetic Qev-train release. No endorsement by NVIDIA or the source authors is implied.
+
+The complete research training mixture and distillation data are not distributed. Other source datasets retain their own terms.
 
 The small JSONL requests in `examples/` were written for Qev and are covered by its Apache-2.0 license. The benchmark result tables and prediction files contain Qev's recorded outputs, not the original benchmark question text.

adapter/adapter_config.json

@@ -32,18 +32,18 @@
   "rank_pattern": {},
   "revision": "68c46c4b3498877f3ef123c856ecfde50c39f404",
   "target_modules": [
+    "down_proj",
+    "in_proj_b",
     "out_proj",
-    "k_proj",
     "v_proj",
-    "down_proj",
-    "gate_proj",
+    "k_proj",
+    "in_proj_z",
+    "o_proj",
     "in_proj_qkv",
-    "in_proj_a",
-    "in_proj_b",
-    "up_proj",
     "q_proj",
-    "in_proj_z",
-    "o_proj"
+    "in_proj_a",
+    "gate_proj",
+    "up_proj"
   ],
   "target_parameters": null,
   "task_type": "FEATURE_EXTRACTION",

adapter/adapter_model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:54314a19f8f9c45fce24f04e14c3615fb1b43be13e231b51a060e9f88d4c57f4
+oid sha256:9554dd117e73c4f8ee34805f9c3223fa1a7d9e35ea050e45116600ecbc5f5217
 size 692518608

benchmarks.json

@@ -1,5 +1,5 @@
 {
-  "date": "2026-09-28",
+  "date": "2026-09-30",
   "metric": "argmax_accuracy",
   "models": {
     "jev": {
@@ -13,11 +13,11 @@
       "source": "JevBench locally rerun; other suites from author reports"
     },
     "qev_9b": {
+      "repo_id": "AustinFu/Qev-9B",
+      "revision": "v0.2.0",
       "base": "Qwen/Qwen3.5-9B-Base",
       "base_revision": "68c46c4b3498877f3ef123c856ecfde50c39f404",
-      "research_name": "BranchKev",
-      "run": "c21-science-wk-late1783-9b-4gpu-r64-late50x3-s17",
-      "step": 2327,
+      "step": 2643,
       "dtype": "bf16",
       "execution": "reference",
       "seed": 17,
@@ -26,6 +26,21 @@
     "qwen35_9b_base": {
       "description": "Frozen native LM head; zero-shot candidate code probabilities",
       "dtype": "bf16"
+    },
+    "qev_2b": {
+      "base": "Qwen/Qwen3.5-2B-Base",
+      "base_revision": "b1485b2fa6dfa1287294f269f5fb618e03d52d7c",
+      "dtype": "bf16",
+      "description": "Probability and representation-response distillation from Qev-9B v0.1.0",
+      "seed": 17,
+      "continuation_steps": 800,
+      "teacher_revision": "v0.1.0"
+    },
+    "qwen35_2b_base": {
+      "description": "Frozen native LM head; zero-shot candidate code probabilities",
+      "dtype": "bf16",
+      "base": "Qwen/Qwen3.5-2B-Base",
+      "base_revision": "b1485b2fa6dfa1287294f269f5fb618e03d52d7c"
     }
   },
   "results": {
@@ -41,12 +56,20 @@
           "accuracy": 0.8780653950953679
         },
         "qev_9b": {
-          "correct": 1296,
-          "accuracy": 0.8828337874659401
+          "correct": 1295,
+          "accuracy": 0.8821525885558583
         },
         "qwen35_9b_base": {
           "correct": 1112,
           "accuracy": 0.7574931880108992
+        },
+        "qev_2b": {
+          "correct": 1266,
+          "accuracy": 0.8623978201634878
+        },
+        "qwen35_2b_base": {
+          "correct": 927,
+          "accuracy": 0.6314713896457765
         }
       }
     },
@@ -68,6 +91,14 @@
         "qwen35_9b_base": {
           "correct": 982,
           "accuracy": 0.7768987341772152
+        },
+        "qev_2b": {
+          "correct": 1079,
+          "accuracy": 0.8536392405063291
+        },
+        "qwen35_2b_base": {
+          "correct": 827,
+          "accuracy": 0.6542721518987342
         }
       }
     },
@@ -89,6 +120,14 @@
         "qwen35_9b_base": {
           "correct": 561,
           "accuracy": 0.7342931937172775
+        },
+        "qev_2b": {
+          "correct": 584,
+          "accuracy": 0.7643979057591623
+        },
+        "qwen35_2b_base": {
+          "correct": 493,
+          "accuracy": 0.6452879581151832
         }
       }
     },
@@ -110,6 +149,14 @@
         "qwen35_9b_base": {
           "correct": 488,
           "accuracy": 0.7439024390243902
+        },
+        "qev_2b": {
+          "correct": 507,
+          "accuracy": 0.7728658536585366
+        },
+        "qwen35_2b_base": {
+          "correct": 427,
+          "accuracy": 0.6509146341463414
         }
       }
     },
@@ -125,12 +172,20 @@
           "accuracy": 0.511
         },
         "qev_9b": {
-          "correct": 546,
-          "accuracy": 0.546
+          "correct": 574,
+          "accuracy": 0.574
         },
         "qwen35_9b_base": {
           "correct": 504,
           "accuracy": 0.504
+        },
+        "qev_2b": {
+          "correct": 387,
+          "accuracy": 0.387
+        },
+        "qwen35_2b_base": {
+          "correct": 312,
+          "accuracy": 0.312
         }
       }
     },
@@ -146,12 +201,20 @@
           "accuracy": 0.9097222222222222
         },
         "qev_9b": {
-          "correct": 135,
-          "accuracy": 0.9375
+          "correct": 134,
+          "accuracy": 0.9305555555555556
         },
         "qwen35_9b_base": {
           "correct": 130,
           "accuracy": 0.9027777777777778
+        },
+        "qev_2b": {
+          "correct": 119,
+          "accuracy": 0.8263888888888888
+        },
+        "qwen35_2b_base": {
+          "correct": 92,
+          "accuracy": 0.6388888888888888
         }
       }
     },
@@ -167,12 +230,20 @@
           "accuracy": 0.7548682703321878
         },
         "qev_9b": {
-          "correct": 631,
-          "accuracy": 0.722794959908362
+          "correct": 620,
+          "accuracy": 0.7101947308132875
         },
         "qwen35_9b_base": {
           "correct": 601,
           "accuracy": 0.6884306987399771
+        },
+        "qev_2b": {
+          "correct": 628,
+          "accuracy": 0.7193585337915235
+        },
+        "qwen35_2b_base": {
+          "correct": 470,
+          "accuracy": 0.5383734249713631
         }
       }
     },
@@ -188,12 +259,20 @@
           "accuracy": 0.703125
         },
         "qev_9b": {
-          "correct": 186,
-          "accuracy": 0.7265625
+          "correct": 182,
+          "accuracy": 0.7109375
         },
         "qwen35_9b_base": {
           "correct": 174,
           "accuracy": 0.6796875
+        },
+        "qev_2b": {
+          "correct": 173,
+          "accuracy": 0.67578125
+        },
+        "qwen35_2b_base": {
+          "correct": 129,
+          "accuracy": 0.50390625
         }
       }
     },
@@ -209,12 +288,20 @@
           "accuracy": 0.7575757575757576
         },
         "qev_9b": {
-          "correct": 188,
-          "accuracy": 0.8138528138528138
+          "correct": 192,
+          "accuracy": 0.8311688311688312
         },
         "qwen35_9b_base": {
           "correct": 175,
           "accuracy": 0.7575757575757576
+        },
+        "qev_2b": {
+          "correct": 172,
+          "accuracy": 0.7445887445887446
+        },
+        "qwen35_2b_base": {
+          "correct": 146,
+          "accuracy": 0.6320346320346321
         }
       }
     }
@@ -225,8 +312,10 @@
       "correct": {
         "jev": 71,
         "kev_9b": 65,
-        "qev_9b": 65,
-        "qwen35_9b_base": 59
+        "qev_9b": 68,
+        "qwen35_9b_base": 59,
+        "qev_2b": 64,
+        "qwen35_2b_base": 44
       }
     },
     "easy": {
@@ -235,7 +324,9 @@
         "jev": 48,
         "kev_9b": 48,
         "qev_9b": 48,
-        "qwen35_9b_base": 48
+        "qwen35_9b_base": 48,
+        "qev_2b": 48,
+        "qwen35_2b_base": 48
       }
     },
     "hard": {
@@ -243,8 +334,10 @@
       "correct": {
         "jev": 79,
         "kev_9b": 62,
-        "qev_9b": 75,
-        "qwen35_9b_base": 68
+        "qev_9b": 76,
+        "qwen35_9b_base": 68,
+        "qev_2b": 60,
+        "qwen35_2b_base": 54
       }
     }
   },
@@ -253,7 +346,7 @@
     "MMLU-Pro: Kev answered 992/1000; its 8 unanswered items count as wrong.",
     "SemIf: the 144 handwritten questions, excluding the 108 perturbations.",
     "Cross-model precisions and execution implementations differ.",
-    "The Qev release is one seed, and these comparisons do not isolate architecture gains."
-  ],
-  "research_report_sha256": "6f094e2e35217089b1ef9365525d3ddbfa46d1354e2faff4b4a25e1bbcfa63d9"
+    "The Qev release is one seed, and these comparisons do not isolate architecture gains.",
+    "The 2B columns use the same clean development and 144-question SemIf subsets as the 9B comparison."
+  ]
 }

evaluation.json

@@ -0,0 +1,86 @@
+{
+  "model": "Qev-9B",
+  "revision": "v0.2.0",
+  "full_suite": {
+    "decision_dev": {
+      "correct": 1295,
+      "total": 1468,
+      "accuracy": 0.8821525885558583,
+      "rejected": 0
+    },
+    "transfer_dev": {
+      "correct": 630,
+      "total": 764,
+      "accuracy": 0.824607329842932,
+      "rejected": 0
+    },
+    "mmlupro": {
+      "correct": 574,
+      "total": 1000,
+      "accuracy": 0.574,
+      "rejected": 0
+    },
+    "semif": {
+      "correct": 237,
+      "total": 252,
+      "accuracy": 0.9404761904761905,
+      "rejected": 0
+    },
+    "scienthoon": {
+      "correct": 620,
+      "total": 873,
+      "accuracy": 0.7101947308132875,
+      "rejected": 0
+    },
+    "wanli": {
+      "correct": 182,
+      "total": 256,
+      "accuracy": 0.7109375,
+      "rejected": 0
+    },
+    "public": {
+      "correct": 192,
+      "total": 231,
+      "accuracy": 0.8311688311688312,
+      "rejected": 0
+    }
+  },
+  "matched_subsets": {
+    "decision_dev_all": {
+      "correct": 1295,
+      "total": 1468
+    },
+    "decision_dev_clean": {
+      "correct": 1105,
+      "total": 1264
+    },
+    "transfer_dev_all": {
+      "correct": 630,
+      "total": 764
+    },
+    "transfer_dev_clean": {
+      "correct": 551,
+      "total": 656
+    },
+    "mmlupro": {
+      "correct": 574,
+      "total": 1000
+    },
+    "semif_handwritten": {
+      "correct": 134,
+      "total": 144
+    },
+    "scienthoon": {
+      "correct": 620,
+      "total": 873
+    },
+    "wanli": {
+      "correct": 182,
+      "total": 256
+    },
+    "jevbench_public": {
+      "correct": 192,
+      "total": 231
+    }
+  }
+}

head.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:6640fcc2045f21e0a0edb4d80d797c2134d160cebba662885cde04efeaafb91c
+oid sha256:61460922e8e45f8a1be335e2845553cbcd8c2228f209654ea67087b2ef45c035
 size 11080748

joint.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:6020d9fde51efc8f12f089fbe10b5f738c00ddf59c8df738ca0eb1edb344a979
+oid sha256:8e9b05c40c52c311c98ce5d2b851acb38fe567e7d6f98c64aada18c8a24b65a9
 size 136

licenses/Crafter-MIT.txt

@@ -0,0 +1,17 @@
+Copyright (c) 2021 Danijar Hafner
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

licenses/HelpSteer3-CC-BY-4.0.txt

@@ -0,0 +1,156 @@
+Creative Commons Attribution 4.0 International
+
+ Creative Commons Corporation (“Creative Commons”) is not a law firm and does not provide legal services or legal advice. Distribution of Creative Commons public licenses does not create a lawyer-client or other relationship. Creative Commons makes its licenses and related information available on an “as-is” basis. Creative Commons gives no warranties regarding its licenses, any material licensed under their terms and conditions, or any related information. Creative Commons disclaims all liability for damages resulting from their use to the fullest extent possible.
+
+Using Creative Commons Public Licenses
+
+Creative Commons public licenses provide a standard set of terms and conditions that creators and other rights holders may use to share original works of authorship and other material subject to copyright and certain other rights specified in the public license below. The following considerations are for informational purposes only, are not exhaustive, and do not form part of our licenses.
+
+Considerations for licensors: Our public licenses are intended for use by those authorized to give the public permission to use material in ways otherwise restricted by copyright and certain other rights. Our licenses are irrevocable. Licensors should read and understand the terms and conditions of the license they choose before applying it. Licensors should also secure all rights necessary before applying our licenses so that the public can reuse the material as expected. Licensors should clearly mark any material not subject to the license. This includes other CC-licensed material, or material used under an exception or limitation to copyright. More considerations for licensors.
+
+Considerations for the public: By using one of our public licenses, a licensor grants the public permission to use the licensed material under specified terms and conditions. If the licensor’s permission is not necessary for any reason–for example, because of any applicable exception or limitation to copyright–then that use is not regulated by the license. Our licenses grant only permissions under copyright and certain other rights that a licensor has authority to grant. Use of the licensed material may still be restricted for other reasons, including because others have copyright or other rights in the material. A licensor may make special requests, such as asking that all changes be marked or described. Although not required by our licenses, you are encouraged to respect those requests where reasonable. More considerations for the public.
+
+Creative Commons Attribution 4.0 International Public License
+
+By exercising the Licensed Rights (defined below), You accept and agree to be bound by the terms and conditions of this Creative Commons Attribution 4.0 International Public License ("Public License"). To the extent this Public License may be interpreted as a contract, You are granted the Licensed Rights in consideration of Your acceptance of these terms and conditions, and the Licensor grants You such rights in consideration of benefits the Licensor receives from making the Licensed Material available under these terms and conditions.
+
+Section 1 – Definitions.
+
+     a.	Adapted Material means material subject to Copyright and Similar Rights that is derived from or based upon the Licensed Material and in which the Licensed Material is translated, altered, arranged, transformed, or otherwise modified in a manner requiring permission under the Copyright and Similar Rights held by the Licensor. For purposes of this Public License, where the Licensed Material is a musical work, performance, or sound recording, Adapted Material is always produced where the Licensed Material is synched in timed relation with a moving image.
+
+     b.	Adapter's License means the license You apply to Your Copyright and Similar Rights in Your contributions to Adapted Material in accordance with the terms and conditions of this Public License.
+
+     c.	Copyright and Similar Rights means copyright and/or similar rights closely related to copyright including, without limitation, performance, broadcast, sound recording, and Sui Generis Database Rights, without regard to how the rights are labeled or categorized. For purposes of this Public License, the rights specified in Section 2(b)(1)-(2) are not Copyright and Similar Rights.
+
+     d.	Effective Technological Measures means those measures that, in the absence of proper authority, may not be circumvented under laws fulfilling obligations under Article 11 of the WIPO Copyright Treaty adopted on December 20, 1996, and/or similar international agreements.
+
+     e.	Exceptions and Limitations means fair use, fair dealing, and/or any other exception or limitation to Copyright and Similar Rights that applies to Your use of the Licensed Material.
+
+     f.	Licensed Material means the artistic or literary work, database, or other material to which the Licensor applied this Public License.
+
+     g.	Licensed Rights means the rights granted to You subject to the terms and conditions of this Public License, which are limited to all Copyright and Similar Rights that apply to Your use of the Licensed Material and that the Licensor has authority to license.
+
+     h.	Licensor means the individual(s) or entity(ies) granting rights under this Public License.
+
+     i.	Share means to provide material to the public by any means or process that requires permission under the Licensed Rights, such as reproduction, public display, public performance, distribution, dissemination, communication, or importation, and to make material available to the public including in ways that members of the public may access the material from a place and at a time individually chosen by them.
+
+     j.	Sui Generis Database Rights means rights other than copyright resulting from Directive 96/9/EC of the European Parliament and of the Council of 11 March 1996 on the legal protection of databases, as amended and/or succeeded, as well as other essentially equivalent rights anywhere in the world.
+
+     k.	You means the individual or entity exercising the Licensed Rights under this Public License. Your has a corresponding meaning.
+
+Section 2 – Scope.
+
+     a.	License grant.
+
+          1. Subject to the terms and conditions of this Public License, the Licensor hereby grants You a worldwide, royalty-free, non-sublicensable, non-exclusive, irrevocable license to exercise the Licensed Rights in the Licensed Material to:
+
+               A. reproduce and Share the Licensed Material, in whole or in part; and
+
+               B. produce, reproduce, and Share Adapted Material.
+
+          2. Exceptions and Limitations. For the avoidance of doubt, where Exceptions and Limitations apply to Your use, this Public License does not apply, and You do not need to comply with its terms and conditions.
+
+          3. Term. The term of this Public License is specified in Section 6(a).
+
+          4. Media and formats; technical modifications allowed. The Licensor authorizes You to exercise the Licensed Rights in all media and formats whether now known or hereafter created, and to make technical modifications necessary to do so. The Licensor waives and/or agrees not to assert any right or authority to forbid You from making technical modifications necessary to exercise the Licensed Rights, including technical modifications necessary to circumvent Effective Technological Measures. For purposes of this Public License, simply making modifications authorized by this Section 2(a)(4) never produces Adapted Material.
+
+          5. Downstream recipients.
+
+               A. Offer from the Licensor – Licensed Material. Every recipient of the Licensed Material automatically receives an offer from the Licensor to exercise the Licensed Rights under the terms and conditions of this Public License.
+
+               B. No downstream restrictions. You may not offer or impose any additional or different terms or conditions on, or apply any Effective Technological Measures to, the Licensed Material if doing so restricts exercise of the Licensed Rights by any recipient of the Licensed Material.
+
+          6.  No endorsement. Nothing in this Public License constitutes or may be construed as permission to assert or imply that You are, or that Your use of the Licensed Material is, connected with, or sponsored, endorsed, or granted official status by, the Licensor or others designated to receive attribution as provided in Section 3(a)(1)(A)(i).
+
+b. Other rights.
+
+          1. Moral rights, such as the right of integrity, are not licensed under this Public License, nor are publicity, privacy, and/or other similar personality rights; however, to the extent possible, the Licensor waives and/or agrees not to assert any such rights held by the Licensor to the limited extent necessary to allow You to exercise the Licensed Rights, but not otherwise.
+
+          2. Patent and trademark rights are not licensed under this Public License.
+
+          3. To the extent possible, the Licensor waives any right to collect royalties from You for the exercise of the Licensed Rights, whether directly or through a collecting society under any voluntary or waivable statutory or compulsory licensing scheme. In all other cases the Licensor expressly reserves any right to collect such royalties.
+
+Section 3 – License Conditions.
+
+Your exercise of the Licensed Rights is expressly made subject to the following conditions.
+
+     a.	Attribution.
+
+          1. If You Share the Licensed Material (including in modified form), You must:
+
+               A. retain the following if it is supplied by the Licensor with the Licensed Material:
+
+                    i. identification of the creator(s) of the Licensed Material and any others designated to receive attribution, in any reasonable manner requested by the Licensor (including by pseudonym if designated);
+
+                    ii. a copyright notice;
+
+                    iii. a notice that refers to this Public License;
+
+                    iv.	a notice that refers to the disclaimer of warranties;
+
+                    v. a URI or hyperlink to the Licensed Material to the extent reasonably practicable;
+
+               B. indicate if You modified the Licensed Material and retain an indication of any previous modifications; and
+
+               C. indicate the Licensed Material is licensed under this Public License, and include the text of, or the URI or hyperlink to, this Public License.
+
+          2. You may satisfy the conditions in Section 3(a)(1) in any reasonable manner based on the medium, means, and context in which You Share the Licensed Material. For example, it may be reasonable to satisfy the conditions by providing a URI or hyperlink to a resource that includes the required information.
+
+          3. If requested by the Licensor, You must remove any of the information required by Section 3(a)(1)(A) to the extent reasonably practicable.
+
+          4. If You Share Adapted Material You produce, the Adapter's License You apply must not prevent recipients of the Adapted Material from complying with this Public License.
+
+Section 4 – Sui Generis Database Rights.
+
+Where the Licensed Rights include Sui Generis Database Rights that apply to Your use of the Licensed Material:
+
+     a.	for the avoidance of doubt, Section 2(a)(1) grants You the right to extract, reuse, reproduce, and Share all or a substantial portion of the contents of the database;
+
+     b.	if You include all or a substantial portion of the database contents in a database in which You have Sui Generis Database Rights, then the database in which You have Sui Generis Database Rights (but not its individual contents) is Adapted Material; and
+
+     c.	You must comply with the conditions in Section 3(a) if You Share all or a substantial portion of the contents of the database.
+For the avoidance of doubt, this Section 4 supplements and does not replace Your obligations under this Public License where the Licensed Rights include other Copyright and Similar Rights.
+
+Section 5 – Disclaimer of Warranties and Limitation of Liability.
+
+     a.	Unless otherwise separately undertaken by the Licensor, to the extent possible, the Licensor offers the Licensed Material as-is and as-available, and makes no representations or warranties of any kind concerning the Licensed Material, whether express, implied, statutory, or other. This includes, without limitation, warranties of title, merchantability, fitness for a particular purpose, non-infringement, absence of latent or other defects, accuracy, or the presence or absence of errors, whether or not known or discoverable. Where disclaimers of warranties are not allowed in full or in part, this disclaimer may not apply to You.
+
+     b.	To the extent possible, in no event will the Licensor be liable to You on any legal theory (including, without limitation, negligence) or otherwise for any direct, special, indirect, incidental, consequential, punitive, exemplary, or other losses, costs, expenses, or damages arising out of this Public License or use of the Licensed Material, even if the Licensor has been advised of the possibility of such losses, costs, expenses, or damages. Where a limitation of liability is not allowed in full or in part, this limitation may not apply to You.
+
+     c.	The disclaimer of warranties and limitation of liability provided above shall be interpreted in a manner that, to the extent possible, most closely approximates an absolute disclaimer and waiver of all liability.
+
+Section 6 – Term and Termination.
+
+     a.	This Public License applies for the term of the Copyright and Similar Rights licensed here. However, if You fail to comply with this Public License, then Your rights under this Public License terminate automatically.
+
+     b.	Where Your right to use the Licensed Material has terminated under Section 6(a), it reinstates:
+
+          1. automatically as of the date the violation is cured, provided it is cured within 30 days of Your discovery of the violation; or
+
+          2. upon express reinstatement by the Licensor.
+
+     c.	For the avoidance of doubt, this Section 6(b) does not affect any right the Licensor may have to seek remedies for Your violations of this Public License.
+
+     d.	For the avoidance of doubt, the Licensor may also offer the Licensed Material under separate terms or conditions or stop distributing the Licensed Material at any time; however, doing so will not terminate this Public License.
+
+     e.	Sections 1, 5, 6, 7, and 8 survive termination of this Public License.
+
+Section 7 – Other Terms and Conditions.
+
+     a.	The Licensor shall not be bound by any additional or different terms or conditions communicated by You unless expressly agreed.
+
+     b.	Any arrangements, understandings, or agreements regarding the Licensed Material not stated herein are separate from and independent of the terms and conditions of this Public License.
+
+Section 8 – Interpretation.
+
+     a.	For the avoidance of doubt, this Public License does not, and shall not be interpreted to, reduce, limit, restrict, or impose conditions on any use of the Licensed Material that could lawfully be made without permission under this Public License.
+
+     b.	To the extent possible, if any provision of this Public License is deemed unenforceable, it shall be automatically reformed to the minimum extent necessary to make it enforceable. If the provision cannot be reformed, it shall be severed from this Public License without affecting the enforceability of the remaining terms and conditions.
+
+     c.	No term or condition of this Public License will be waived and no failure to comply consented to unless expressly agreed to by the Licensor.
+
+     d.	Nothing in this Public License constitutes or may be interpreted as a limitation upon, or waiver of, any privileges and immunities that apply to the Licensor or You, including from the legal processes of any jurisdiction or authority.
+
+Creative Commons is not a party to its public licenses. Notwithstanding, Creative Commons may elect to apply one of its public licenses to material it publishes and in those instances will be considered the “Licensor.” Except for the limited purpose of indicating that material is shared under a Creative Commons public license or as otherwise permitted by the Creative Commons policies published at creativecommons.org/policies, Creative Commons does not authorize the use of the trademark “Creative Commons” or any other trademark or logo of Creative Commons without its prior written consent including, without limitation, in connection with any unauthorized modifications to any of its public licenses or any other arrangements, understandings, or agreements concerning use of licensed material. For the avoidance of doubt, this paragraph does not form part of the public licenses.
+
+Creative Commons may be contacted at creativecommons.org.

provenance.json

@@ -0,0 +1,77 @@
+{
+  "checkpoint": {
+    "repo_id": "AustinFu/Qev-9B",
+    "revision": "v0.2.0",
+    "base": "Qwen/Qwen3.5-9B-Base",
+    "base_revision": "68c46c4b3498877f3ef123c856ecfde50c39f404",
+    "step": 2643,
+    "dtype": "bf16",
+    "execution": "reference",
+    "seed": 17,
+    "temperature": 1.0
+  },
+  "config": "configs/qev-9b.json",
+  "training_recipe": {
+    "main_records": 39605,
+    "main_questions": 45887,
+    "principle_records": 4459,
+    "boundary_records": 600,
+    "boundary_partition": "main",
+    "late_records": 1783,
+    "late_fraction": 0.5,
+    "late_repeats": 3,
+    "epochs": 2,
+    "global_batch": 32
+  },
+  "protocol": {
+    "execution": "full causal reference",
+    "metric": "argmax accuracy; rejected questions count as wrong",
+    "jevbench_version": "v1.4.2",
+    "jevbench_upstream_revision": "1df665e3956d7aab7fa0208ff6c4f2d8557f9f90",
+    "trained_limits": {
+      "max_state": 1024,
+      "max_question": 512,
+      "max_candidate": 256,
+      "max_path": 2048,
+      "max_candidates": 128
+    }
+  },
+  "results": {
+    "decision_dev_all": {
+      "correct": 1295,
+      "total": 1468
+    },
+    "decision_dev_clean": {
+      "correct": 1105,
+      "total": 1264
+    },
+    "transfer_dev_all": {
+      "correct": 630,
+      "total": 764
+    },
+    "transfer_dev_clean": {
+      "correct": 551,
+      "total": 656
+    },
+    "mmlupro": {
+      "correct": 574,
+      "total": 1000
+    },
+    "semif_handwritten": {
+      "correct": 134,
+      "total": 144
+    },
+    "scienthoon": {
+      "correct": 620,
+      "total": 873
+    },
+    "wanli": {
+      "correct": 182,
+      "total": 256
+    },
+    "jevbench_public": {
+      "correct": 192,
+      "total": 231
+    }
+  }
+}

training_config.json

@@ -43,7 +43,8 @@
     "none_insert_prob": 0.2,
     "none_insert_absent_frac": 0.5,
     "none_insert_exempt_sources": [
-      "jev_distill/"
+      "jev_distill/",
+      "synthetic/hs3_preference_boundary/"
     ],
     "late_split": "late_train",
     "late_fraction": 0.5,