更换unsloth grpo的训练数据集并验证

src/train_model_grpo.py

@@ -209,7 +209,7 @@ class ModelTrainer:
                 correctness_reward_func,
             ],
             args = training_args,
-            train_dataset = dataset,
+            train_dataset = train_dataset,
         )
         
         # 训练模型