respect dataset adapter

eval_protocol/pytest/evaluation_test.py

@@ -211,7 +211,7 @@ def evaluation_test(
     completion_params = parse_ep_completion_params_overwrite(completion_params)
     original_completion_params = completion_params
     passed_threshold = parse_ep_passed_threshold(passed_threshold)
-    data_loaders = parse_ep_dataloaders(data_loaders)
+    data_loaders = parse_ep_dataloaders(data_loaders, dataset_adapter=dataset_adapter)
     custom_invocation_id = os.environ.get("EP_INVOCATION_ID", None)
 
     # ignore other data input params when dataloader is provided

eval_protocol/pytest/evaluation_test_utils.py

@@ -21,6 +21,7 @@
     EvaluationThresholdDict,
     Status,
 )
+from eval_protocol.common_utils import load_jsonl
 from eval_protocol.data_loader import DynamicDataLoader
 from eval_protocol.data_loader.models import EvaluationDataLoader
 from eval_protocol.pytest.rollout_processor import RolloutProcessor
@@ -288,10 +289,21 @@ def _rows_from_jsonl(path: str) -> list[EvaluationRow]:
 
 def parse_ep_dataloaders(
     dataloaders: Sequence[EvaluationDataLoader] | EvaluationDataLoader | None,
+    *,
+    dataset_adapter: Callable[[list[dict[str, Any]]], list[EvaluationRow]] | None = None,
 ) -> Sequence[EvaluationDataLoader] | EvaluationDataLoader | None:
+    """When ``EP_JSONL_PATH`` is set, load JSONL as raw dicts and run ``dataset_adapter`` if provided.
+
+    Without ``dataset_adapter``, rows are built with ``EvaluationRow(**dict)`` (legacy behavior),
+    which skips custom label fields that adapters normally attach.
+    """
     try:
         load_from_jsonl_path = os.getenv("EP_JSONL_PATH")
         if load_from_jsonl_path:
+            if dataset_adapter is not None:
+                return DynamicDataLoader(
+                    generators=[lambda path=load_from_jsonl_path, da=dataset_adapter: da(load_jsonl(path))]
+                )
             return DynamicDataLoader(generators=[lambda path=load_from_jsonl_path: _rows_from_jsonl(path)])
     except Exception:
         pass