fix(inspect): align GenerateConfig field names with current inspect-ai

- rename CLI/eval_set kwarg 'frequence_penalty' -> 'frequency_penalty' (typo; inspect-ai field name)
- rename 'log_probs' -> 'logprobs' (inspect-ai field name)
- pass 'response_format' only when the installed inspect-ai still accepts it, warn otherwise (removed in >= 0.3.141)

Fixes #1327: a fresh 'pip install lighteval' resolves inspect-ai to the latest release and 'lighteval eval' crashed at startup with a pydantic ValidationError. Verified on inspect-ai 0.3.257: the repro command from the issue now completes.

Author: sk8ordie84

src/lighteval/main_inspect.py

@@ -43,6 +43,25 @@
 
 logger = logging.getLogger(__name__)
 
+def _response_format_kwargs(response_format: str | None) -> dict:
+    """inspect-ai removed GenerateConfig.response_format (>= 0.3.141).
+
+    Pass the flag through only when the installed inspect-ai accepts it, so a
+    fresh install does not crash at startup with a pydantic ValidationError.
+    """
+    if response_format is None:
+        return {}
+    from inspect_ai.model import GenerateConfig
+
+    if "response_format" in GenerateConfig.model_fields:
+        return {"response_format": response_format}
+    logger.warning(
+        "--response-format is not supported by the installed inspect-ai (%s); ignoring it.",
+        getattr(__import__("inspect_ai"), "__version__", "unknown"),
+    )
+    return {}
+
+
 
 @task
 def get_inspect_ai_task(
@@ -267,7 +286,7 @@ def eval(  # noqa C901
             rich_help_panel=HELP_PANEL_NAME_1,
         ),
     ] = None,
-    frequence_penalty: Annotated[
+    frequency_penalty: Annotated[
         float | None,
         Option(
             help="Number between -2.0 and 2.0, Penalizes tokens that appear in the text too frequently, reducing repetition.",
@@ -309,7 +328,7 @@ def eval(  # noqa C901
             rich_help_panel=HELP_PANEL_NAME_1,
         ),
     ] = None,
-    log_probs: Annotated[
+    logprobs: Annotated[
         bool | None,
         Option(
             help="Returns log probabilities for each token in the generated text", rich_help_panel=HELP_PANEL_NAME_1
@@ -474,19 +493,19 @@ def eval(  # noqa C901
         temperature=temperature,
         top_p=top_p,
         top_k=top_k,
-        frequence_penalty=frequence_penalty,
+        frequency_penalty=frequency_penalty,
         presence_penalty=presence_penalty,
         seed=seed,
         stop_seqs=stop_seqs,
         num_choices=num_choices,
         best_of=best_of,
-        log_probs=log_probs,
+        logprobs=logprobs,
         top_logprobs=top_logprobs,
         cache_prompt=cache_prompt,
         reasoning_effort=reasoning_effort,
         reasoning_tokens=reasoning_tokens,
         reasoning_history=reasoning_history,
-        response_format=response_format,
+        **_response_format_kwargs(response_format),
         parallel_tool_calls=parallel_tool_calls,
         max_tool_output=max_tool_output,
         internal_tools=internal_tools,