[misc] fix packing and eval plot (#7623)

src/llamafactory/train/sft/workflow.py

@@ -110,7 +110,15 @@ def run_sft(
         trainer.save_metrics("train", train_result.metrics)
         trainer.save_state()
         if trainer.is_world_process_zero() and finetuning_args.plot_loss:
-            plot_loss(training_args.output_dir, keys=["loss", "eval_loss", "eval_accuracy"])
+            keys = ["loss"]
+            if isinstance(dataset_module["eval_dataset"], dict):
+                keys += sum(
+                    [[f"eval_{key}_loss", f"eval_{key}_accuracy"] for key in dataset_module["eval_dataset"].keys()], []
+                )
+            else:
+                keys += ["eval_loss", "eval_accuracy"]
+
+            plot_loss(training_args.output_dir, keys=keys)
 
     if training_args.predict_with_generate:
         tokenizer.padding_side = "left"  # use left-padding in generation