facebookresearch / fairseq-lua

Facebook AI Research Sequence-to-Sequence Toolkit
Other
3.74k stars 616 forks source link

Contributing pretrained models (wmt17.zh-en) #97

Closed twairball closed 7 years ago

twairball commented 7 years ago

Hi guys,

I'd like to contribute my models i've trained on wmt17 news commentary for chinese-english.

Are there any guidelines on dataset? I randomly split dataset 80% training 10% validation 10% test.

(website mentions they don't have preprocessed training or dev sets for zh-en) http://www.statmt.org/wmt17/translation-task.html

Model Epochs Training Time BLEU4 (beam1) BLEU4 (beam5) BLEU4 (beam10) BLEU4 (beam20)
fconv 25 ~2hrs 63.49 62.22 62.52 62.74
blstm 30 ~8hrs 64.59 64.15 64.38 63.76
convenc 47 ~7hrs 50.91 56.71 56.83 53.66

(Comment edited with links to models and training results)

michaelauli commented 7 years ago

Thanks! Before adding the model, please add a README.md similar to the other pretrained models (https://github.com/facebookresearch/fairseq#pre-trained-models) describing the data pre-processing. Each archive also contains source/target dictionaries as well as an optional bpe code file (if the data was bpe encoded), e.g.,

$ ls wmt16.en-ro.fconv-float bpecodes dict.en.th7 dict.ro.th7 model.th7 README.md

twairball commented 7 years ago

Thanks for the reply michael -- I'll fix that up soon.

neoyang0620 commented 4 years ago

Hi, the link contains pretrained-models doesn't work. Could you please give me another valid link?