Greg Tatum
Greg Tatum
In [Ludicrously Fast Neural Machine Translation](https://aclanthology.org/D19-5632.pdf), they test a variety of decoder configurations for faster models. In #174 @eu9ene showed that a larger decoder helps improve the COMET score for...
We will need to retrain models that don't have the robustness fixes from using OpusTrainer. - [ ] bg-en - [ ] de-en - [ ] en-bg - [ ]...
Short sentences are frequently removed from parallel datasets, so there aren't enough to train on. In HPLT 2.0 the data is filtered at the document level, rather than sentence level....
It's not guaranteed that parallel "sentences" are actually sentences. We could write a script to detect how many sentences are in each parallel datum, and then attempt to extract them...
We already statistically generate word alignment information, it should be possible to go through parallel datasets, and generate word pairs of the most common words that are aligned. Since the...
This will let me get a profile of it.
I'm working on some performance optimization to bring down our CI times, and I found some tasks are taking 12 minutes to resolve due to issues downloading fetches and artifacts....
We had an incident where the uk-en model broke in Nightly. We should automate tests before release so that it catches issues like. We should run each release channel of...
It costs money to store models in the cloud. We could save a bit, and make the output of the train tasks a bit less confusing if we just stored...