answer cannot be None from LM

Author: EYH0602

src/tfbench/lm/_anthropic.py

@@ -1,6 +1,6 @@
 import os
 from anthropic import Anthropic
-from ._types import LM, LMAnswer, ReasoningEffort, EFFORT_TOKEN_MAP
+from ._types import LM, LMAnswer, ReasoningEffort, EFFORT_TOKEN_MAP, NoneResponseError
 from .settings import MAX_TOKENS
 
 CLAUDE_MODELS = [
@@ -52,7 +52,7 @@ def _gen(self, prompt: str) -> LMAnswer:
                 break
 
         if text is None:
-            raise ValueError("No text content returned from the model.")
+            raise NoneResponseError(self.model_name)
 
         return LMAnswer(answer=text)
 
@@ -97,7 +97,7 @@ def _gen(self, prompt: str) -> LMAnswer:
             if text and thinking:
                 break
 
-        if text is None or thinking is None:
-            raise ValueError("No text content returned from the model.")
+        if text is None:
+            raise NoneResponseError(self.model_name)
 
         return LMAnswer(answer=text, reasoning_steps=thinking)

src/tfbench/lm/_google.py

@@ -6,7 +6,7 @@
 from google import genai
 from google.genai.types import GenerateContentConfig, ThinkingConfig
 
-from ._types import LM, LMAnswer, ReasoningEffort, EFFORT_TOKEN_MAP
+from ._types import LM, LMAnswer, ReasoningEffort, EFFORT_TOKEN_MAP, NoneResponseError
 
 GEMINI_MODELS = [
     "gemini-2.0-flash",
@@ -61,7 +61,10 @@ def _gen(self, prompt: str) -> LMAnswer:
                 system_instruction=[self.instruction],
             ),
         )
-        return LMAnswer(answer=response.text)
+        content = response.text
+        if content is None:
+            raise NoneResponseError(self.model_name)
+        return LMAnswer(answer=content)
 
 
 class GeminiReasoning(LM):
@@ -105,7 +108,7 @@ def _gen(self, prompt: str) -> LMAnswer:
 
         candidate = response.candidates[0]
         if not candidate.content or not candidate.content.parts:
-            raise ValueError("No content in the candidate response.")
+            raise NoneResponseError(self.model_name)
 
         answer = ""
         thinking = ""

src/tfbench/lm/_openai.py

@@ -3,7 +3,7 @@
 from openai import OpenAI, NOT_GIVEN
 from openai.types.shared.reasoning_effort import ReasoningEffort
 
-from ._types import LM, LMAnswer
+from ._types import LM, LMAnswer, NoneResponseError
 
 OAI_MODELS = [
     "gpt-3.5-turbo-0125",
@@ -57,8 +57,12 @@ def _gen(self, prompt: str) -> LMAnswer:
             model=self.model_name,
         )
 
+        content = completion.choices[0].message.content
+        if content is None:
+            raise NoneResponseError(self.model_name)
+
         return LMAnswer(
-            answer=completion.choices[0].message.content,
+            answer=content,
         )
 
 

src/tfbench/lm/_types.py

@@ -14,7 +14,7 @@
 
 @dataclass
 class LMAnswer:
-    answer: str | None
+    answer: str
     reasoning_steps: str | None = None
 
 
@@ -46,3 +46,8 @@ def _gen(self, prompt) -> LMAnswer:
     "high": 8192,
     "off": 0,
 }
+
+
+class NoneResponseError(Exception):
+    def __init__(self, model: str):
+        super().__init__(f"None response from model: {model}")

src/tfbench/lm/_vllm.py

@@ -4,15 +4,7 @@
 from openai import OpenAI
 
 from ..env import ENV
-from ._types import LM, LMAnswer
-
-
-_CHAT_TEMPLATE = """# Instructions
-{instruction}
-
-## Task
-{task}
-"""
+from ._types import LM, LMAnswer, NoneResponseError
 
 
 class VLLMChat(LM):
@@ -64,8 +56,12 @@ def _gen(self, prompt: str) -> LMAnswer:
         )
 
         message = completion.choices[0].message
+        content = message.content
+        if content is None:
+            raise NoneResponseError(self.model_name)
+
         return LMAnswer(
-            answer=message.content,
+            answer=content,
             reasoning_steps=(
                 message.reasoning_content  # type: ignore
                 if hasattr(message, "reasoning_content")