[Feat] convert sft dataset in chat template style

Author: realtmxi

data/generate_sft_verl.py

@@ -0,0 +1,40 @@
+import os
+import argparse
+from datasets import Dataset, load_dataset
+from tqdm import tqdm
+
+
+def make_map_fn(split):
+    def process_fn(example, idx):
+        return {
+            "data_source": "openmanus-rl",
+            "prompt": example['conversations'],
+            "ability": "instruction-following",
+            "reward_model": {
+                "style": "none",
+                "ground_truth": None
+            },
+            "extra_info": {
+                "split": split,
+                "index": idx,
+                "id": example['id']
+            }
+        }
+    return process_fn
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--output_dir', required=True, help="Output directory for processed parquet")
+    parser.add_argument('--split', type=str, default="train")
+
+    args = parser.parse_args()
+
+    # Load from Hugging Face Hub
+    dataset = load_dataset("CharlieDreemur/OpenManus-RL", split=args.split)
+
+    # Apply mapping to Verl format
+    dataset = dataset.map(function=make_map_fn(args.split), with_indices=True)
+
+    os.makedirs(args.output_dir, exist_ok=True)
+    dataset.to_parquet(os.path.join(args.output_dir, f"{args.split}.parquet"))