Deepseek R1 cutting off

I’ve been testing with Deepseek R1 recently and I’ve noticed that it generates very long chain of thought reasoning on many questions, simple or not. The issue here is that it seems that there’s some sort of safety measure made to stop the model from generating overly long responses, meaning that I usually don’t even get the full answer/generation. I could probably implement an automatic continuation system into my application but was wondering if this was intentional or not.

1 Like

@vibright10 interesting let me tinker with it and ask engineering.

-Coby

Thanks, @coby.adams!
It’s probably a limit on the max tokens per generation since it seems to cut off at around the same amount of tokens every time.