Low accuracy for Qwen-2.5-7b-instruct model

**Describe the bug**
I have tested finetuning with qwen2.5-7b-instruct model.
with the scripts:
python3 examples/llm/finetune/automodel.py

All with default values,
8 H100 GPUS
Batch size=1
The training completed in less than 4 hours.

But accuracy of the model seems to be very low.
Did anyone face this issue?




