Fix QA example

[Rocketknight1]

The QA example assumes the model has a CLS token, but many newer models don't. This PR falls back to the BOS token if one isn't present, and finally falls back to the start of the sequence if that doesn't work either.

Fixes #30570

[amyeroberts]

Thanks for the quick fix on this!

Just a request to do a quick test run on the old default models, and then with one of the newer ones on just one of these scripts to make sure it all runs smoothly

[Rocketknight1]

Will do now before merging!

[Rocketknight1]

Tried running it on Falcon myself, but it turns out there are multiple other issues here - in particular, the code also assumes a PAD token is present. Fundamentally, this example just isn't suited to fine-tuning a CLM like Falcon!

I think we can maybe still merge the PR (since there are likely MLM models either now or in the future that will be missing CLS), but @ananegru you'll probably need to use a masked language model as a base rather than a causal language model! Maybe one of the DeBERTa models?

[Rocketknight1]

Just confirmed that the example still works with this branch and microsoft/deberta-v3-base

My bad for going through all this before realizing the root of the problems was a CLM being used in MLM examples, though!

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[amyeroberts]

My bad for going through all this before realizing the root of the problems was a CLM being used in MLM examples, though!

I didn't notice either! Happy for this to be merged in if it still works with existing models. Thanks again for adapting the scripts so quickly