Allow multilabel classification mode for widgets in the models repo

[avidale]

🚀 Feature request

1. Enable multilabel classification mode and regression mode for the widgets in the model repo.
2. Create the corresponding tags that can be read from the model card.

Motivation

Models for sequence classification by default support three modes: binary/multiclass classification, multilabel classification, and regression. However, the widgets in the model repository support only multiclass mode (where probabilities of classes sum to 1). This can be misleading for the users who taste the models using the widgets. For example, my model https://huggingface.co/cointegrated/rubert-tiny-toxicity, which is intended for multilabel classification, but the widget normalizes the predicted probabilities to sum to 1, which leads to confisuion of the potential users of the model.

Your contribution

If you show me where to start, I could start working on implementing this feature. However, currently I don't know what part of the Huggingface repository is responsible for widgets and underlying computations.

[mishig25]

Hi @avidale, To start contributing widgets, checkout the widgets folder in this repo.

In terms of contribution, create a branch and start pushing your updates there. Moreover, you can create draft PR so that we can discuss specific details etc as you make progress. You can find existing PRs for widgets here where you can see the overall contribution process.

Please let me know of any other questions.