After SFT, the model make repetitions

[cnlinxi]

Hello, I tried to use the following two datasets:

• OpenThoughts-114k
• OpenR1-Math220k

After SFT on Qwen2.5-Math-7B, the model with packing=true had serious auto-regression, and the repetitions with packing=false was reduced. I have fixed SFT model make repetitions during the inference phase.

Any ideas?

[qgallouedec]

Maybe this? huggingface/trl#3091