Output reasoning_text

flexeval/core/evaluate_chat_response.py

@@ -229,6 +229,8 @@ def evaluate_chat_response(  # noqa: C901, PLR0912
         }
         if output["lm_output"].raw_text:
             restructured_output["raw_lm_output"] = output["lm_output"].raw_text
+        if output["lm_output"].reasoning_text:
+            restructured_output["reasoning_text"] = output["lm_output"].reasoning_text
         restructured_outputs.append(restructured_output)
 
     return metrics_summary_dict, restructured_outputs

tests/core/test_evaluate_chat_response.py

@@ -66,6 +66,8 @@ def test_evaluate_chat_response(
     # Therefore, in any case the system message should be in the first turn.
     assert outputs[0]["extra_info"]["messages"][0]["role"] == "system"
 
+    assert outputs[0]["reasoning_text"] == "reasoning_text"
+
     if use_tools:
         assert isinstance(outputs[0]["extra_info"]["tool_calls"], list)
         assert isinstance(outputs[0]["extra_info"]["tools"], list)

tests/dummy_modules/lm.py

@@ -45,6 +45,7 @@ def _batch_generate_chat_response(
         return [
             LMOutput(
                 text=f"This is response to `{messages[-1]['content']}` with kwargs {kwargs}",
+                reasoning_text="reasoning_text",
                 finish_reason="length",
                 tool_calls=tc,
                 tool_call_validation_result="CompleteToolCall" if tc else "TextOnly",