The Kaggle competitions have closed, but you can still see how a submission would have scored on the public leaderboard.
What gets scored
- BalAcc@10, the competition metric: top-10 balanced accuracy over the 50-word vocabulary (chance 0.2).
- BalAcc@1: the same, counting only the top prediction.
- OVMI: out-of-vocabulary mutual information, in bits.
The file is the one you would have uploaded to Kaggle: an index column and one score column per vocabulary word, with a row for every test index of the track.
Submitting from the command line
Handy for large Broad files or scripted evaluation. The response is JSON with the scores.
gzip -k submission.csv curl --data-binary @submission.csv.gz \ "https://eval.libribrain.com/api/submissions?track=deep"