[Update] Minor updates

opencompass/datasets/TheoremQA/number_utils.py

@@ -48,7 +48,7 @@ def clean_units(pred_str: str):
 
 
 def number_it(num):
-    from latex2sympy2 import latex2sympy
+    from latex2sympy2_extended import latex2sympy
     if isinstance(num, (int, float)):
         return num
 

opencompass/datasets/TheoremQA/utils.py

@@ -17,7 +17,7 @@ def time_limit(seconds: float):
 
 
 def extract_theoremqa_answer(pred: str, answer_flag: bool = True):
-    from latex2sympy2 import latex2sympy
+    from latex2sympy2_extended import latex2sympy
 
     if any([option in pred.lower() for option in ['yes', 'true']]):
         pred = 'True'

opencompass/models/claude_sdk_api.py

@@ -33,6 +33,7 @@ class ClaudeSDK(BaseAPIModel):
         max_seq_len: int = 2048,
         meta_template: Optional[Dict] = None,
         temperature: Optional[float] = 0.0,
+        thinking: Optional[Dict] = None,
         retry: int = 2,
     ):
         super().__init__(path=path,
@@ -49,6 +50,7 @@ class ClaudeSDK(BaseAPIModel):
         self.anthropic = Anthropic(api_key=key)
         self.model = path
         self.temperature = temperature
+        self.thinking = thinking
 
     def generate(
         self,
@@ -108,11 +110,26 @@ class ClaudeSDK(BaseAPIModel):
         while num_retries < self.retry:
             self.wait()
             try:
-                responses = self.anthropic.messages.create(
+                api_params = {
-                    model=self.model,
+                    'model': self.model,
-                    max_tokens=max_out_len,
+                    'max_tokens': max_out_len,
-                    temperature=self.temperature,
+                    'temperature': self.temperature,
-                    messages=messages)
+                    'messages': messages,
+                }
+
+                if self.thinking is not None:
+                    api_params['thinking'] = self.thinking
+                    api_params['stream'] = True
+
+                responses = self.anthropic.messages.create(**api_params)
+
+                # Handle new response format
+                for content in responses.content:
+                    if content.type == 'text':
+                        return content.text
+
+                # If no text type content is found, return the first
+                # content (backward compatibility)
                 return responses.content[0].text
             except Exception as e:
                 self.logger.error(e)

opencompass/models/openai_api.py

@@ -652,7 +652,6 @@ class OpenAISDK(OpenAI):
                     self.logger.info('Start calling OpenAI API')
                 responses = self.openai_client.chat.completions.create(
                     **query_data, timeout=timeout)  # timeout in seconds
-
                 if self.verbose:
                     self.logger.info(
                         'Successfully get response from OpenAI API')
@@ -660,10 +659,18 @@ class OpenAISDK(OpenAI):
                         self.logger.info(responses)
                     except Exception:
                         pass  # noqa F841
-                if not responses.choices:
+
+                # Check if response is empty or content is empty
+                if not responses.choices or not responses.choices[
+                        0].message.content:
                     self.logger.error(
-                        'Response is empty, it is an internal server error \
+                        'API response is empty, it might be due to excessive '
-                            from the API provider.')
+                        'input length or an internal server error '
+                        'from your API provider.')
+                    num_retries += 1
+                    # Continue to retry instead of returning empty response
+                    continue
+
                 return responses.choices[0].message.content
 
             except (BadRequestError, APIStatusError) as e:

requirements/extra.txt

@@ -11,12 +11,10 @@ faiss_gpu==1.7.2
 -e git+https://github.com/open-compass/human-eval.git#egg=human-eval
 # IFEval
 langdetect
-# TheoremQA
-latex2sympy2==1.9.1
 # Lawbench, leval
 ltp
 # Math
-math-verify
+math-verify[antlr4_11_0]
 # Taco, apps Dataset
 pyext
 # Law Bench