Open kirill-fedyanin opened 11 months ago
Also, wonder why we switching from max_new_tokens to just max_tokens here
Also, wonder why we switching from max_new_tokens to just max_tokens here
Probably to comply with openAI API format? @cant-access-rediska0123 can you comment on that?
I was running the polygraph_eval with this example config https://github.com/IINemo/lm-polygraph/blob/main/examples/configs/polygraph_eval_wmt14_ende.yaml
I got a warning about too long string
It didn't failed but i almost sure it would mess up with the results
My wild guess is that stat calculators max_lengths in not connected to max_generated_tokens in ue manager itself. But i didn't really look into it for now
I'm pretty sure this is due to this line here: https://github.com/IINemo/lm-polygraph/blob/main/src/lm_polygraph/utils/model.py#L166
For some we reason we set a very low limit on the input sequence length, which triggers the warning during generation. It doesn't fail though, because in reality the model (Llama2 for example) is capable of handling much longer sequences than measly 256 tokens.
Not sure why weird even set this value, probably should remove it, along with all other code for loading and configuring whitebox models.
@rvashurin check please
I was running the polygraph_eval with this example config https://github.com/IINemo/lm-polygraph/blob/main/examples/configs/polygraph_eval_wmt14_ende.yaml
I got a warning about too long string
It didn't failed but i almost sure it would mess up with the results![Screenshot from 2023-10-09 11-11-07](https://github.com/IINemo/lm-polygraph/assets/2440125/263dc315-7516-46f1-a2b9-a33be921667b)
My wild guess is that stat calculators max_lengths in not connected to max_generated_tokens in ue manager itself. But i didn't really look into it for now