换用github jwjohns/unsloth-GRPO-qwen2.5 验证GRPO训练模型

src/qwen_notebook_clone.py

@@ -69,9 +69,11 @@ def extract_hash_answer(text: str) -> str | None:
 
 # Dataset preparation
 from datasets import load_dataset, Dataset
+from modelscope.msdatasets import MsDataset
 
 def get_gsm8k_questions(split="train") -> Dataset:
-    data = load_dataset('openai/gsm8k', 'main')[split]
+    # data = load_dataset('openai/gsm8k', 'main')[split]
+    data =  MsDataset.load('openai-mirror/gsm8k', subset_name='main', split=split)
     data = data.map(lambda x: {
         'prompt': [
             {'role': 'system', 'content': SYSTEM_PROMPT},