The model is described in Multilingual Translation with Extensible Multilingual Pretraining and Finetuning
Thank you for your interest. We are doing a few clean-up and will release them shortly.
Waiting for the model too :)
Thanks
@tangyuq Is the benchmark included in Beyond English-Centric Multilingual Machine Translation?
We'll be adding a full evaluation on the ML50 benchmark in an updated version. The main reason why I haven't added it yet is that we want to update ML50's evaluation pairs done on WMT20 to the WMT20 test sets. When ML50 was originally released, the WMT20 test sets were not released yet. So we'll update both works together.
A portion of ML50 evaluation, from the original mBART paper, is available in Beyond English Centric MMT work already.
Most helpful comment
We'll be adding a full evaluation on the ML50 benchmark in an updated version. The main reason why I haven't added it yet is that we want to update ML50's evaluation pairs done on WMT20 to the WMT20 test sets. When ML50 was originally released, the WMT20 test sets were not released yet. So we'll update both works together.
A portion of ML50 evaluation, from the original mBART paper, is available in Beyond English Centric MMT work already.