Greg Tatum
Greg Tatum
We're getting into languages that aren't in the original Flores datasets, and need to update to Flores+ https://huggingface.co/datasets/openlanguagedata/flores_plus It's gated behind a user agreement on Hugging Face, so we need...
It's not shippable as-is. We could just distill a base-memory and see if it's a good enough quality, but the teacher is pretty low too. It might be worth auditing...
Teacher: 86.14, -3.87% Student: 81.42, -9.89% https://mozilla.github.io/translations/model-registry/?searchString=Bosnian&showModels=true&score=vs-google I believe this is #681, since Bosnian is digraphic for Cyrillic and Latin. We should be able to train a `bs-en` model just...
Teacher: [87.65](https://firefox-ci-tc.services.mozilla.com/tasks/BZFB3MA4QWuK4oWEafn2vQ) -1.39% Student: [85.52](https://firefox-ci-tc.services.mozilla.com/tasks/RPJvxpa4SJSi_Ybrh0lHAg) -3.92% The simplest here would be to distill a `base-memory`. I'm wondering if we should do split vocabs as well, but we would probably have...
The distillation gap is quite large, and the teacher is OK. We should distill as base-memory to make it shippable. Teacher: [86.01](https://firefox-ci-tc.services.mozilla.com/tasks/c_SPjl3PTqa3OIsm5YnxDg) -2.23% Student: [82.88](https://firefox-ci-tc.services.mozilla.com/tasks/YWJnykjpSuqNpWjAW9xyqg) -6.09%
`no` is a macro language tag for `nn` and `nb`. `no-en` is technically a multilingual model, while `en-no` I went ahead and trained as I'm guessing it's similar to `en->pt`...
In the Spring 2024 run, [we trained a teacher](https://wandb.ai/moz-translations/en-vi/workspace?nw=nwusergtatum), but the student was never distilled because the COMET score was a bit low, 87.52 -2.59%. The training curves for the...
For some reason when training `en-af` I had some infrastructure issues around getting the downloads. I don't know if it was some bug in my corpus continuation code or if...
This is currently blocked because Tagalog and Filipino both have distinct language codes (tl, and fil respectively). However, for our initial models I don't think we will need to distinguish...
I just hit this once, and I haven't tested if this is an intermittent issue or not. ``` old_e = 3, n_items = 856819, dynamic = 1 eflomal: eflomal.c:394: text_alignment_sample:...