| { |
| "config": { |
| "run_name": "9b-chat-v1", |
| "description": "The v1 recipe on the CHAT checkpoint Qwen/Qwen3.5-9B (thinking off). 4b-chat-v1 scored 72.49 (best 4B ever, +2.2 over 4b-v1) with no synthetic data, so the base checkpoint is the lever. Read against 9b-v1 (73.29): the candidate for the published 9B v2.", |
| "base_model": "Qwen/Qwen3.5-9B", |
| "base_revision": "c202236235762e1c871ad0ccb60c8ee5ba337b9a", |
| "data_dir": "/workspace/jeb/data-v1", |
| "dataset_path": "/workspace/jeb/data-v1/train.jsonl", |
| "eval_dataset_path": "/workspace/jeb/data-v1/calib.jsonl", |
| "output_dir": "/workspace/jeb/runs/9b-chat-v1", |
| "method": "lora", |
| "num_epochs": 1, |
| "batch_size": 8, |
| "gradient_accumulation_steps": 1, |
| "learning_rate": 0.0001, |
| "lr_scheduler_type": "cosine", |
| "lora_rank": 16, |
| "lora_alpha": 32, |
| "max_seq_length": 2048, |
| "warmup_steps": 30, |
| "weight_decay": 0.0, |
| "max_grad_norm": 1.0, |
| "use_gradient_checkpointing": true, |
| "attn_implementation": "sdpa", |
| "eval_steps": 200, |
| "logging_steps": 10, |
| "seed": 17, |
| "decide": { |
| "objective": "candidate_ce", |
| "shuffle_choice_options": true, |
| "lora_dropout": 0.05, |
| "calib_eval_limit": 358, |
| "target_modules": "all-linear", |
| "score_targets": "ordinal", |
| "score_ordinal_adjacent": 0.2 |
| }, |
| "prompt_source_sha256": "d2660ebec28bd3f1704235bda88d24a397c1c62475e740519cb8ef2d08f25fdd", |
| "prompt_source_commit": "e5c089386e0239c9eb270eeb490d181722b8da5b", |
| "sweep": { |
| "letter": "h", |
| "smoke": false, |
| "eval_only": false, |
| "max_steps": -1, |
| "eval_limit": 0, |
| "eval_repeats": 5, |
| "eval_repeats_for": "jevals-=5,nimble-eval=3,kev-transfer=3,typed-decisions=2,kev-decision=1,nimble-public=1", |
| "eval_batch_size": 16, |
| "eval_latency_sample": 30, |
| "temperature_target": "train", |
| "notes": [] |
| } |
| }, |
| "target_modules": "all-linear", |
| "lora_modules": [ |
| "down_proj", |
| "gate_proj", |
| "in_proj_a", |
| "in_proj_b", |
| "in_proj_qkv", |
| "in_proj_z", |
| "k_proj", |
| "o_proj", |
| "out_proj", |
| "q_proj", |
| "up_proj", |
| "v_proj" |
| ], |
| "trainable_params": 43278336, |
| "prompt_source_commit": "e5c089386e0239c9eb270eeb490d181722b8da5b", |
| "prompt_source_sha256": "d2660ebec28bd3f1704235bda88d24a397c1c62475e740519cb8ef2d08f25fdd", |
| "chat_template_sha256": "a4aee8afcf2e0711942cf848899be66016f8d14a889ff9ede07bca099c28f715", |
| "single_token_labels": 68, |
| "chat_template_kwargs": { |
| "add_generation_prompt": true, |
| "enable_thinking": false, |
| "thinking": false |
| }, |
| "train_examples": 14892, |
| "calib_examples": 689, |
| "truncated_prompts": 6, |
| "steps": 1862, |
| "wall_clock_s": 5333.4, |
| "load_s": 50.6, |
| "peak_memory_allocated_gb": 31.59, |
| "final_calib": { |
| "calib_accuracy": 0.8603351955307262, |
| "calib_nll": 0.4381291315885571, |
| "calib_accuracy_choice": 0.875, |
| "calib_nll_choice": 0.4329145725041313, |
| "calib_accuracy_noul": 0.8490566037735849, |
| "calib_nll_noul": 0.34073033509579664, |
| "calib_accuracy_score": 0.8548387096774194, |
| "calib_nll_score": 0.5267721637743564 |
| }, |
| "hardware": "NVIDIA A100 80GB PCIe", |
| "torch": "2.11.0+cu128", |
| "adapter_dir": "/workspace/jeb/runs/9b-chat-v1/adapter", |
| "fp32_candidate_logits": true, |
| "score_targets": "ordinal", |
| "score_ordinal_adjacent": 0.2 |
| } |
