Update utils.py

[ZihanWang314]

update compute_accuracy to deal with the cases where str_chosen and str_rej got the same scores, which is probably what the developers don't want

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint.

[lvwerra]

Thanks for the PR! Can you elaborate a bit more why this is bad and what bad side effects this has?

[ZihanWang314]

When a model gives exactly same scores for the sentence to accept and the sentence to reject, the current evaluation metric will set this pair to be "correct". However, the above scenario is usually caused by the model having some problems in encoding sentences (e.g. over-quantized), or the two sentences exactly being the same, which we both don't expect. I think putting a warning here and defining this pair as "wrong" or "random" would be helpful for the developers to observe these problems.

[lvwerra]

I see, thanks! In that case I'd prefer to just add a warning but leave the computation logic the same. Does that make sense?

[ZihanWang314]

sure it makes sense. Thanks for your suggestions!