Heavily improve automatic model card generation + Patch XLM-R

[tomaarsen]

Hello!

Pull Request overview

• Heavily improve automatic model card generation

Example generated output

These model cards were fully generated - none of this was edited except for one sentence: See [train.py](train.py) for the training script.:

• tomaarsen/span-marker_bert-base-uncased-keyphrase-inspec: no evaluation during training
• tomaarsen/span-marker-bert-base-uncased-bionlp: with evaluation during training

Details

Add the following sections to the README body:

• Model Description (encoder model, maximum sequence length, maximum entity length, training dataset, language, license)
• Model Sources (repo link, thesis link)
• Model labels (a table)
• Direct uses (inference, a code snippet)
• Downstream use (finetuning, a code snippet)
• Training set metrics (sentence length, entities per sentence)
• Training hyperparameters
• Training results
• Environmental Impact (emissions, training time)
• Training hardware
• Framework versions

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 (f1, recall, precision)
• base_model

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

TODO

• Tests
• Additional documentation

cc: @davanstrien I think this might really interest you. Perhaps you have some useful feedback!

• Tom Aarsen