The <eos> token randomly pops out during the inference, making the text generation stops early.

[BadisG]

Hello,

Several people have reported having this issue for the v1.1 model, which wasn't a bug for the v1.0 one. Do you plan on fixing that in the future?

[lan2720]

same question, anyone knows how to stop the generation?

[Thireus]

Same issue. <\s> randomly shows up from time to time in the middle of the assistant's answer.

[merrymercy]

I also found this several times. I guess it is a problem with our training data but I haven't found the reason. Could you try some pre or post-processing as a workaround for now?