For helix, I notice that GitHub is listed as a data source, but there's nothing in the docs about this. I'd really love to see what a model trained on my commonly used git repos (which generally are newer than The Stack etc), and in particular their commit history. Ideally these would make it easier for code completion to have the historical context as well as the current code to play with in determining what to write next.

I often wonder how you'd go about organizing training data for a full historic github repo in a way that makes sense for training (or RAG)? The vast majority of the data is previous changes to the repo. I think this would generally mean that it would outweigh the current information and cause problems (i.e. old method names before refactoring etc.)

Also, perhaps being able to expand that out to doing the same thing for a bunch of consumers of the library that I'm maintaining would be neat.

Sprinkle in the PR and Issue history, docs website, API docs, and discord history and I think you'd have a helluva model.