fixes

src/axolotl/utils/callbacks.py

@@ -217,7 +217,9 @@ def bench_eval_callback_factory(trainer, tokenizer):
             metrics: Dict[str, float],  # pylint: disable=unused-argument
             **kwargs,  # pylint: disable=unused-argument
         ):
-            data_loader = trainer.get_bench_dataloader(bench_dataset)
+            data_loader = trainer.get_bench_dataloader(
+                bench_dataset.remove_columns(["input", "subject", "output"])
+            )
             trainer.model.eval()
             preds, refs = [], []
             loss_bench = 0

src/axolotl/utils/trainer.py

@@ -687,7 +687,7 @@ def setup_trainer(cfg, train_dataset, eval_dataset, model, tokenizer, total_num_
             return_tensors="pt",
             **data_collator_kwargs,
         ),
-        bench_data_collat0r=transformers.DataCollatorForSeq2Seq(
+        bench_data_collator=transformers.DataCollatorForSeq2Seq(
             tokenizer,
             return_tensors="pt",
             **data_collator_kwargs,