Handle potentially long sequences with DataCollatorForCompletionOnlyLM

[tannonk]

This PR resolves #643 and provides a patch in DataCollatorForCompletionOnlyLM for handling long sequences which may or may not contain a valid response_template or a valid instruction_template.

For problematic instances where no response_template or instruction_template is found, we set the labels to the ignore_idx. As a result, problematic instances are ignored in the loss computation, but still allow for training to continue.

[lvwerra]

Looks good to me, thanks for fixing! Just a small nit then we can merge.

[HuggingFaceDocBuilderDev]

The documentation is not available anymore as the PR was closed or merged.

[lvwerra]

Also could you run the code quality tests with make precommit? That should fix the CI.

[tannonk]

Hey @lvwerra, thanks for the review. The requested changes have been made and make precommit passes on my end.

[lvwerra]

Looks like the tests are not passing, yet.

[younesbelkada]

Hi @tannonk
Thanks a lot for the PR!
In my opinion, currently raising a RuntimeError really helps for understanding whether things worked correctly or not for users, I am afraid the warning is not a strong enough indicator to users. Maybe it is better to keep it and add a flag allow_ignore_not_matched for advanced users and keep the previous behaviour untouched, what do you think? cc @lvwerra - if that's not a good idea I am happy to merge the PR as it is as well

[younesbelkada]

Discussed with @lvwerra , indeed if the sequence is too long the training would completely break and this could happen in the middle of a training, which is too odd. Let's merge it! Thanks again @tannonk for your work on this!