Greg Tatum
Greg Tatum
We have: ``` pipeline/train/train.sh taskcluster/scripts/pipeline/train-taskcluster.sh taskcluster/scripts/pipeline/train_taskcluster.py ``` It will be much simpler to put this into a single training script. We should probably do this after our first big training...
The models tend to break down when presented with single numbers, or lists of numbers. For example de-en translates ``` 1 2 3 4 5 6 7 8 9 ```...
They have a rather unique behavior, and so it would be helpful to have them separated out. Something like: * `tests/task` * `tests/unit`
In PR #157 I added additional alphabet support. This information is available by professional translators in the CLDR data: https://github.com/unicode-org/cldr-json/blob/0876ec40e13d54c0a6b6456392802d4de7e059cb/cldr-json/cldr-misc-full/main/sl/characters.json It would be nice to consume that JSON and automate...
Similar to #736, we should discard sentences that translate the domain suffix of a website, like Amazon.com -> Amazon.it With a regex such as `/[a-z]+\.com\b/` we could identify a URL...
As a follow-up to #481, we need support for the "Q" field, which is Quarters.
In #771 I ran an experiment to see the effects of the size of the distillation corpus for the change in the COMET score for the students. Adding more data...
Training a second teacher improves performance only slightly. It may be more cost efficient to take the quality hit and remove it. Comet Change | Average Type -- | --...
An experiment for #231 da-en is one of our best models from the spring-2024 run. The teacher ensemble had a COMET score of 0.9013. The student COMET was 0.8950, with...
In Firefox the memory size of the inference engine is quite large in wasm. There aren't good memory tools to analyze the wasm. Instead, we should compile it natively, and...