Copy rl/scripts from HuggingFaceBio/FineDNA-RL@2315ddc

rl/scripts/create_dataset.py

@@ -0,0 +1,62 @@
+# /// script
+# dependencies = ["datasets>=4", "huggingface_hub"]
+# ///
+"""Build FineDNA-RL from the Carbon-X GRPO prompt blend.
+
+The blend is a single subset with weight 1.0, so the trainer's mixture loader shuffles it twice with the mixture seed
+(once per subset, once after concatenation). This script does the same and labels each row with the domain of its
+`task_group`, so the published split holds the rows the trainer sees, in the same order.
+
+    python scripts/create_dataset.py --output_dataset HuggingFaceBio/FineDNA-RL
+"""
+
+import argparse
+
+from datasets import load_dataset
+from huggingface_hub import HfApi
+
+
+SEED = 0
+SOURCE = "HuggingFaceBio/Rollouts-Carbon-14B-SFT-v36.52-step-000019119"
+SOURCE_CONFIG = "blend_all_v0"
+SOURCE_REVISION = "205d8b22b8e29a364b8922eb3dc96728cb6f65d7"
+COLUMNS = ["prompt", "answer_text", "task_group", "task_type", "verifier_metadata", "gold_answer_json"]
+DOMAINS = {
+    "basic_dna": "dna_computation",
+    "dna_skills": "dna_computation",
+    "seqrecqa_position": "dna_computation",
+    "bioreason": "variant_effect_prediction",
+    "vepqa": "variant_effect_prediction",
+    "genomeqa": "sequence_classification",
+    "bio_mcq": "biology_qa",
+    "if": "instruction_following",
+}
+
+
+def build_dataset():
+    ds = load_dataset(SOURCE, SOURCE_CONFIG, split="train", revision=SOURCE_REVISION).select_columns(COLUMNS)
+    ds = ds.add_column("domain", [DOMAINS[group] for group in ds["task_group"]])
+    return ds.shuffle(seed=SEED).shuffle(seed=SEED)
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--output_dataset", default="HuggingFaceBio/FineDNA-RL")
+    args = parser.parse_args()
+
+    dataset = build_dataset()
+    print(dataset)
+    dataset.push_to_hub(
+        args.output_dataset, split="train", private=True, commit_message="Build FineDNA-RL from the Carbon-X GRPO blend"
+    )
+    HfApi().upload_file(
+        path_or_fileobj=__file__,
+        path_in_repo="scripts/create_dataset.py",
+        repo_id=args.output_dataset,
+        repo_type="dataset",
+        commit_message="Publish the script that built this dataset",
+    )
+
+
+if __name__ == "__main__":
+    main()