新增enable_thinking参数以支持推理过程中的思考功能；更新repetition_penalty和max_new_tokens配置以优化数据清理策略。

weclone/core/inference/offline_infer.py

@@ -26,6 +26,7 @@ def vllm_infer(
     max_samples: Optional[int] = None,
     vllm_config: str = "{}",
     save_name: str = "generated_predictions.jsonl",
+    enable_thinking: bool = False,
     temperature: float = 0.95,
     top_p: float = 0.7,
     top_k: int = 50,
@@ -61,6 +62,7 @@ def vllm_infer(
             "top_k": top_k,
             "max_new_tokens": max_new_tokens,
             "repetition_penalty": repetition_penalty,
+            "enable_thinking": enable_thinking,
         }
     )
 

weclone/data/clean/strategies.py

@@ -117,8 +117,9 @@ class LLMCleaningStrategy(CleaningStrategy):
             template=self.make_dataset_config.template,
             temperature=0,
             guided_decoding_class=QaPairScore,
-            repetition_penalty=1.2,
+            repetition_penalty=1.5,
             bad_words=[r"\n"],
+            max_new_tokens=self.make_dataset_config.messages_max_length + 1024,  # add prompt length
         )
 
         parsed_scores: List[QaPairScore] = []