train & eval scripts

[lannelin]

No description provided.

[lannelin]

Hi @eddableheath @J-Dymond . I've run this locally for 2 epochs to train a model which gets zero one acc: 0.23 hamming acc: 0.92 on validation set (though not with seed set :( . )
Let me know if you'd like to review and then run on baskerville - possibly with slightly tweaked hparams - or prefer me/one of you to run this on baskerville first.

[J-Dymond]

Looks good! I think it would be a good idea to train it for longer on Baskerville. This performs better than the zero-shot, as I'm sure we expected, so I can work to get this into the current pipelines. We also have the more difficult levels of topic label which we can use, though we could probably move to training on those labels that at a later stage?

[lannelin]

this has now run on baskerville and trained model available in project dir under distilbert_topic_classifier/.
I've made a couple of small changes in latest commits to fix the readme and to allow for loading only a single split of the data

[lannelin]

closes #23

[J-Dymond]

These changes look good to me, will merge them into inference pipeline branch before I open a pull request.