Refactor training logs & warmup_proportion

[tomaarsen]

Resolves #473

Hello!

Pull Request overview

• Refactor training logs.
• Update how warmup_steps are defined, now uses number of training samples if training samples > args.max_steps

Details

Training logs

Now we display the number of unique pairs (instead of number of examples, which was broken). The rest is only updated slightly.

Update how warmup_steps are defined

It now uses number of training samples if training samples > args.max_steps alongside the warmup_proportion. This should make the warmup_proportion match reality.

Thanks @michael-brunzel for raising this!

• Tom Aarsen

[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.