fixing tests

src/lighteval/logging/evaluation_tracker.py

@@ -523,8 +523,8 @@ def push_results_to_tensorboard(  # noqa: C901
                 else:
                     tb_context.add_scalar(f"{prefix}/{task_name}/{metric}", value, global_step=global_step)
         # e.g. MMLU
-        for name, averages in bench_averages.items():
-            for metric, values in averages.items():
+        for name, values in bench_averages.items():
+            for metric, values in values.items():
                 hlog(f"Pushing average {name} {metric} {sum(values) / len(values)} to tensorboard")
                 tb_context.add_scalar(f"{prefix}/{name}/{metric}", sum(values) / len(values), global_step=global_step)
 

src/lighteval/tasks/lighteval_task.py

@@ -367,6 +367,7 @@ def create_requests_from_tasks(  # noqa: C901
     lm: BaseModel,
     max_samples: int,
     evaluation_tracker: "EvaluationTracker",
+    use_chat_template: bool,
 ) -> Tuple[dict[RequestType, list[Request]], dict[TaskExampleId, Doc]]:
     """
     Takes a task dict and a fewshot dict and returns a dict of requests, a dict of docs, and a dict of requests origins.
@@ -428,6 +429,7 @@ def create_requests_from_tasks(  # noqa: C901
                         max_model_length=lm.max_length,
                         sampler=rnd,
                         tokenizer=lm.tokenizer,
+                        use_chat_template=use_chat_template,
                     )
                     doc.num_effective_few_shots = num_effective_few_shots
                     doc.num_asked_few_shots = num_fewshot