<thinking> vs <think>
#2
by omercelik - opened
I see both in the output, is this a pre-processing error?
This could be, I didn't notice these issues with my minimal testing. Tomorrow I can go back and verify the pipeline to make sure the tags are correct.
Could you please verify that you are using the recommended sampling parameters
Ok I looked into the data and am now certain that there was no tags ever used on my end. The only thing I can think of (if not a sampling or inference level issue) is the model forgot that it uses tags due to some weird masking error in unsloth's latest notebook.