Remove deprecated old_benchmarking utilities

src/art/model.py

@@ -25,7 +25,10 @@
 from .preprocessing.moe_routing import attach_moe_routing_metadata_to_choice
 from .trajectories import Trajectory, TrajectoryGroup
 from .types import TrainSFTConfig
-from .utils.trajectory_logging import write_trajectory_groups_parquet
+from .utils.trajectory_logging import (
+    calculate_step_std_dev,
+    write_trajectory_groups_parquet,
+)
 
 if TYPE_CHECKING:
     from wandb.sdk.wandb_run import Run
@@ -991,11 +994,6 @@ async def log(
                 group_key = f"group_{metric}"
                 averages[group_key] = sum(values) / len(values)
 
-        # Calculate average standard deviation of rewards within groups
-        from .utils.old_benchmarking.calculate_step_metrics import (
-            calculate_step_std_dev,
-        )
-
         averages[reward_std_dev_key] = calculate_step_std_dev(trajectory_groups)
 
         # Merge in any additional metrics passed directly

src/art/utils/benchmarking/aggregate_trajectories.py

@@ -21,8 +21,7 @@ async def load_aggregated_trajectories(
     Load trajectories and aggregate metrics at the step level.
 
     This function builds on top of load_trajectories to provide step-level
-    aggregation similar to load_benchmarked_models, but returns a DataFrame
-    instead of custom objects.
+    aggregation as a DataFrame.
 
     Parameters
     ----------