mirror of
https://github.com/Helsinki-NLP/OPUS-MT-train.git
synced 2025-01-08 10:48:25 +03:00
1.7 KiB
1.7 KiB
Things to do
Bugs
- something is wrong with multi-threaded data preparation
- balancing data for multilingual models does not work well with one lang-pair that is tiny
General settings
- better hyperparameters for low-resource setting (lower batch sizes, smaller vocabularies ...)
- better data selection (data cleaning / filtering); use opus-filter?
- better balance between general data sets and backtranslations
Backtranslation
- status: basically working, need better integration?!
- add backtranslations to training data
- can use monolingual data from tokenized wikipedia dumps: https://sites.google.com/site/rmyeid/projects/polyglot
- https://dumps.wikimedia.org/backup-index.html
- better in JSON: https://dumps.wikimedia.org/other/cirrussearch/current/
Fine-tuning and domain adaptation
- status: basically working
- do we want to publishfine-tuned data or rather the fina-tuning procedures? (using a docker container?)
Show-case some selected language pairs
- collaboration with wikimedia
- focus languages: Tagalog (tl, tgl), Central Bikol (bcl), Malayalam (ml, mal), Bengali (bn, ben), and Mongolian (mn, mon)
Tatoeba MT models
Labels are only taken from test data but this can be a problem if there are relevant data sets that will be missed out
- example: nor (there is only nno nob in the test data but most of the data for Norwegian is only tagged as nor_Latn);
- another example: hbs (hbs labels do not exist in test data)
- possible solution: take all labels from train data; problem: some noisy labels may influence the model a lot and it would be better to leave them out (wrong script data etc); another issue: over-sampling data sets that only exist in train data may damage the model