diff --git a/README.md b/README.md index a19380b..941a264 100644 --- a/README.md +++ b/README.md @@ -78,34 +78,33 @@ Then convert the training files to parquet (consumed by `verl`): python dataprocess/prepare_sft.py \ --task 1 \ --sft_data_path data/train/sft_task1_45961.jsonl \ - --sft_remain_path data/train/sft_raw_pool_52813.jsonl \ + --sft_remain_path data/train/SFT_50K.jsonl \ --local_dir data/abforge_task1_sft python dataprocess/prepare_sft.py \ --task 2 \ --sft_data_path data/train/sft_task2_37019.jsonl \ - --sft_remain_path data/train/sft_raw_pool_52813.jsonl \ + --sft_remain_path data/train/SFT_50K.jsonl \ --local_dir data/abforge_task2_sft # RL python dataprocess/prepare_task1_rl.py \ - --input data/train/rl_task1_30000.jsonl \ + --input data/train/RL_task1_30K.jsonl \ --local_dir data/abforge_task1_rl python dataprocess/prepare_task2_rl.py \ - --input data/train/rl_task2_rubric_v2_30000.jsonl \ + --input data/train/RL_task2_30K.jsonl \ --local_dir data/abforge_task2_rl ``` The held-out evaluation files are under `data/eval/`. Use -`ablationbench_1000_rubric_v2.jsonl` for the full rubric-v2 benchmark and -`ablationbench_200_rubric_v2.jsonl` for the clean 200-instance subset. +`ablationbench_1000.jsonl` for the full benchmark and +`ablationbench_200.jsonl` for the clean 200-instance human-evaluation subset. -> **Task defaults.** Task 1 SFT/RL preprocessing keeps papers with 2–6 GT -> focuses by default. Task 1 RL/eval defaults use the v18.3 count penalty (one -> free extra bullet above GT, soft penalty `0.005` up to 7 bullets, hard penalty -> `0.05` beyond 7). Override with `TASK1_COUNT_*` environment variables. See -> [`configs/task1.md`](configs/task1.md) and [`configs/task2.md`](configs/task2.md). +> **Task defaults.** Task 1 SFT/RL preprocessing keeps papers with 2–6 +> ground-truth focuses by default. See +> [`configs/task1.md`](configs/task1.md) and [`configs/task2.md`](configs/task2.md) +> for the configurable defaults. ## 🛠️ Training @@ -200,10 +199,9 @@ scripts/evaluate_task2.sh outputs/task2_infer.jsonl - `configs/` — task defaults and an environment-variable template (`env.example`). - `examples/` — tiny schema examples only; full data lives on Hugging Face. -> **Notes.** No Slurm scripts, training data, checkpoints, logs, secrets, or -> machine-specific paths are committed. Task 1 currently uses the v18.3 TM+RQ -> formulation: 2–6 bullets, each with a target module and a high-level research -> question. +> **Notes.** No training data, checkpoints, logs, secrets, or machine-specific +> paths are committed. Task 1 produces 2–6 bullets, each pairing a target module +> with a high-level research question. ## 🙏 Acknowledgements diff --git a/configs/task1.md b/configs/task1.md index 59816fd..3e5c251 100644 --- a/configs/task1.md +++ b/configs/task1.md @@ -8,6 +8,6 @@ - Data filter: keep papers with 2-6 GT focuses by default. - Output format: `` with 2-6 bullets. Each bullet contains `- Target Module: ...` and an indented `- Research Question: ...`. -- RL reward/eval count penalty default follows v18.3: - `TASK1_COUNT_FREE_EXTRA=1`, `TASK1_COUNT_SOFT_RATE=0.005`, - `TASK1_COUNT_HARD_THRESHOLD=7`, `TASK1_COUNT_HARD_RATE=0.05`. +- The RL reward and evaluation share an optional count-related penalty that is + fully configurable through `TASK1_COUNT_*` environment variables (see + `reward/task1_candidate_utils.py` for the available knobs and their defaults). diff --git a/dataprocess/prepare_sft.py b/dataprocess/prepare_sft.py index 647d2b3..fcedb6e 100644 --- a/dataprocess/prepare_sft.py +++ b/dataprocess/prepare_sft.py @@ -11,7 +11,7 @@ ablation-plan synthesis responses. Usage: python dataprocess/prepare_sft.py --task 1 \ --sft_data_path data/train/sft_task1_45961.jsonl \ - --sft_remain_path data/train/sft_raw_pool_52813.jsonl \ + --sft_remain_path data/train/SFT_50K.jsonl \ --tokenizer_path Qwen/Qwen3-8B \ --local_dir data/abforge_task1_sft \ --val_size 200 diff --git a/evaluation/eval_task1.py b/evaluation/eval_task1.py index 8dd5cc2..3a5b4f9 100644 --- a/evaluation/eval_task1.py +++ b/evaluation/eval_task1.py @@ -384,7 +384,7 @@ def compute_match_rate(matches: List[Dict], n_gt: int) -> Optional[float]: def compute_count_penalty(n_pred: int, n_gt: int) -> float: - """Use the same GT-relative count penalty family as the v18.3 reward.""" + """Use the same GT-relative count penalty family as the RL reward.""" return compute_count_penalty_v2_from_env(n_pred=n_pred, n_gt=n_gt) diff --git a/reward/task1_candidate_utils.py b/reward/task1_candidate_utils.py index f0fa96b..0b7ad33 100644 --- a/reward/task1_candidate_utils.py +++ b/reward/task1_candidate_utils.py @@ -148,7 +148,7 @@ def compute_count_penalty_v2( cap: float = 0.5, free_extra: int = 1, ) -> float: - """GT-relative count penalty (v18.3). + """GT-relative count penalty. n_pred ≤ n_gt + free_extra: no penalty (free zone). n_gt + free_extra < n_pred ≤ hard_threshold: soft_rate per excess bullet above the free zone. diff --git a/reward/task1_reward.py b/reward/task1_reward.py index 9629a8f..34bd24e 100644 --- a/reward/task1_reward.py +++ b/reward/task1_reward.py @@ -219,7 +219,7 @@ def compute_format_factor(raw_response: str, result_text: str) -> Dict[str, floa def compute_count_penalty(n_pred: int, n_gt: int) -> float: - """GT-relative count penalty (v18.3). Delegates to compute_count_penalty_v2_from_env. + """GT-relative count penalty. Delegates to compute_count_penalty_v2_from_env. n_pred ≤ n_gt + TASK1_COUNT_FREE_EXTRA: no penalty. n_gt + TASK1_COUNT_FREE_EXTRA < n_pred ≤ TASK1_COUNT_HARD_THRESHOLD (default 7):