After fine-tuning, I ran into a problem where the model’s inference results are garbled.

[dadax7]

image_path:"/data/vjuicefs_ai_camera_qgroup_ql/11190414/Rex-Omni-master/images/000042.jpg"
raw_output:"<|object_ref_start|>human<|object_ref_end|><|box_start|><0><289><118>ଉ,<122><303><187>𝔸,<171><329><278>𝙰,<335><338>+F⽗,<354><334>代谢,<412><325><467>Ճ,edia<|box_end|>,<476><311><581>ቨ,<531><323><598><2>"

The inference output is as shown above. The fine-tuning loss decreased normally from around 17 to approximately 3. What could be causing this issue?

[tzjtatata]

I think the sft loss as 3 is still high.

[github-actions]

This issue has been automatically closed because it has been inactive for 30 days.