Heavily improve automatic model card generation

[tomaarsen]

Hello!

Pull Request overview

• Heavily improve automatic model card generation
• Additionally:
  • Always return dicts in id2label/label2id
  • Set default W&B project as "setfit" if not set via ENV var yet
  • Set fsspec<2023.12.0 due to breakages with older datasets versions

Example generated output

These model cards were fully generated - none of this was edited:

• tomaarsen/setfit-all-MiniLM-L6-v2-sst2-32-shot
• tomaarsen/setfit-paraphrase-mpnet-base-v2-sst2-8-shot
• tomaarsen/setfit-absa-bge-small-en-v1.5-restaurants-aspect
• tomaarsen/setfit-absa-bge-small-en-v1.5-restaurants-polarity

Details

Add the following sections to the README body:

• Model Description (embedding model, classification head, maximum sequence length, number of classes, training dataset, language, license)
• Model Sources (repo link, paper link, blogpost link)
• Model labels (a table)
• Evaluation results (accuracy on test set, usually)
• Direct uses (inference, a code snippet)
• Training set metrics (sentence length, training sample count per label)
• Training hyperparameters
• Training results/logs (with load_best_model_at_end checkpoint highlighted)
• Environmental Impact (emissions, training time)
• Training hardware
• Framework versions
• Citation

Beyond that, I added some values to the README metadata:

• language
• license
• library_name
• tags
• datasets
• metrics
• widget
• pipeline_tag
• co2_eq_emissions
• model-index (accuracy)
• base_model
• inference

The values for these are automatically computed via the Trainer and a ModelCardCallback.

TODO

• Basic tests
• Additional stress tests if possible
• Additional documentation

---

• Tom Aarsen

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint.