Add token classification eval with CoNLL 2003

[tylerjthomas9]

Changes

This PR adds support for CoNLL 2003 token classification/entity recognition. It should be easier to integrate other token classification datasets now that the classes have been built out.

Using the overall_f1 metric from seqeval, here are the HF and Mosaic BERT ablations:

• HF BERT: 90.51
• Mosaic BERT: 60.92

I trained a quick checkpoint of Flex BERT and verified that this also ran without errors, and got a score of 64.28.

Here are the

Discussions
I am not aware of any discussions on the topic, but the BertForTokenClassification class was left as TBD.

bert24/src/bert_layers/model.py

Line 669 in 664db03

class BertForTokenClassification(BertPreTrainedModel):

Tests

• Is the new feature tested? (Not always necessary for all changes -- just adding to the checklist to keep track)
• Have you ran all the tests?
• Do the tests all pass?
• If not, have you included an explanation of which tests this PR breaks and/or why (below this checklist)

[bclavie]

Hey! Thanks for adding this. As discussed, we won't be merging outside evals right now (as we finalised the training) runs but we'll be revisiting this shortly after.

Thanks again!