Support all negative values in ClassLabel

[lhoestq]

We usually use -1 to represent a missing label, but we should also support any negative values (some users use -100 for example). This is a regression from datasets 2.3

Fix #4508

[HuggingFaceDocBuilderDev]

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

[mariosasko]

Thanks, LGTM!

[davidbenton]

Thanks for this fix! I'm not sure what the release timeline is, but FYI #4508 is a breaking issue for transformer token classification using Trainer and PyTorch. PyTorch defaults to -100 as the ignored label for negative log loss, so switching labels to -1 leads to index errors using Trainer defaults.

As a workaround, I'm using master branch directly (pip install git+https://github.com/huggingface/datasets.git@master for anyone who needs to do the same) until this gets released.

[lhoestq]

The new release 2.4 fixes the issue, feel free to update datasets :)

pip install -U datasets

[l-uuz]

@lhoestq I hope it's OK to ping you here. I've noticed that encode_example does only work with -1. I already created #7645 to fix the documentation, but then I stumbled across your original changes to the docs text in this PR.

I am talking about this part in ClassLabel -> encode_example:

datasets/src/datasets/features/features.py

Lines 1126 to 1129 in e71b0b1

	# Allowing -1 to mean no label.
	if not -1 <= example_data < self.num_classes:
	raise ValueError(f"Class label {example_data:d} greater than configured num_classes {self.num_classes}")
	return example_data