Open mnoukhov opened 2 years ago
A working branch of FLORES v1
Tested reproduce.sh neen and got 12.5 BLEU on devtest after 2 iterations of BT (compared to README's 15.9). I used 1 RTX8000 and the full pipeline ran in ~100 hours (after adjusting max_tokens to 16000 to eliminate unnecessary update_freq)
reproduce.sh neen
max_tokens
update_freq
Download Issues:
download_indic.sh
en-hi
Other issues:
fairseq-train
min-lr
args
cfg
@guzmanhe thanks! I rebased + merged the two previous PRs so if they are accepted I should have no merge conflicts but let me know if there are issues
A working branch of FLORES v1
Tested
reproduce.sh neen
and got 12.5 BLEU on devtest after 2 iterations of BT (compared to README's 15.9). I used 1 RTX8000 and the full pipeline ran in ~100 hours (after adjustingmax_tokens
to 16000 to eliminate unnecessaryupdate_freq
)Download Issues:
download_indic.sh
git command #22en-hi
dataset as it is now manual download only (https://www.cfilt.iitb.ac.in/iitb_parallel/dataset.html)Other issues:
fairseq-train
args e.g.min-lr
args
to omegaconfcfg