code/Abforge_Training/abforge_release_fixes.patch

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: `<Result>` 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):