Fix: add assistant_only_loss=False to prevent all labels being masked to -100

train.py

@@ -122,6 +122,7 @@ training_args = SFTConfig(
     weight_decay=0.01,
     max_length=2048,
     packing=False,
+    assistant_only_loss=False,  # CRITICAL FIX: prevent all labels being masked to -100
     bf16=True,
     logging_strategy="steps",
     logging_steps=10,
@@ -133,7 +134,7 @@ training_args = SFTConfig(
     push_to_hub=True,
     hub_model_id=output_dir,
     report_to="trackio",
-    run_name=output_dir.replace("/", "-") + "-optimal-resume",
+    run_name=output_dir.replace("/", "-") + "-optimal",
     remove_unused_columns=False,
     disable_tqdm=True,
     dataset_num_proc=4,