Hi, when training NLLB200 model we used a bit more data that what can be found here.
In particular there was a number of languages where we felt we already had enough bitext in CC-Matrix and just reused that.

This PR allows to download directions from http://data.statmt.org/cc-matrix/ when they aren't found on AllenAI published website.

The PR is ready but I believe I also need to update the datasets_info.json which I haven't done yet.

I think I'm done, I've updated the datasets_info.json file.
To make it clearer that I've only added data and not modified any existing data, I've put the new metadata on a new line.

Allen Institute for AI org

@gwenzek Thanks for the PR. I tried it out, and it works as expected. We should also update the README section on languages (https://huggingface.co/datasets/allenai/nllb#languages), so that it is easier for people to find the ccmatrix pairs. The pair-mapping distinction in ccmatrix_lang_pairs is definitely cleaner, but we might want to have the combined version simply for the ease of lookup.

what about I move the content of cc_matrix_lang_pairs.py at the bottom of nllb_lang_pairs.py ?
And I would convert it to contains the name with NLLB200 convention.

akshitab changed pull request status to open

Hi, thanks for the review. I believe I need to update the metadata/hashsum, so please wait a bit before merging.

ready to merge :-) I have tested downloading several dataset and it seems to work fine :-)

akshitab changed pull request status to merged

Sign up or log in to comment