Joel Barry

Hi @Simsso - no, I had no response on this. This was a while ago. I think I ended up just using the original code instead, at https://github.com/tmikolov/word2vec.git

This might be relevant: http://www.gossamer-threads.com/lists/lucene/java-user/186647

For fine-tuning the many-to-one model, I've found that I additionally need to use `--arch mbart_large` in the `fairseq_train` command with `--task translation_multi_simple_epoch`. The current docs say to use `--arch transformer`...

@Remorax - not sure this will help you, but here are some more specifics of what I did. Here is the commit I used: ``` $ git log | head...

@Remorax I just confirmed I can still run my fine-tuning script. I also double checked the model I started with and it is the same as yours: https://dl.fbaipublicfiles.com/fairseq/models/mbart50/mbart50.ft.n1.tar.gz (I re-downloaded...

@Remorax I think @gegallego is right on point. If I don't pass --lang-dict: ``` fairseq-train $data_dir --save-dir $save_dir \ --finetune-from-model $mbart_checkpoint \ --encoder-normalize-before --decoder-normalize-before \ --arch mbart_large --layernorm-embedding \ --task...