That may be the issue. I can get about 6k lines of data today and in time, could have a bunch more. Each line represents a full request (user input, model output, full prompt, etc.) so I would intuitively think that 2k of these is good enough.

And it's actually good enough for llama-2-7b, insofar as it transforms this model from "terrible with my prompt in production today" to "this could be workable, actually". But my bar is to be better than gpt-3.5 and require less text in my prompt, but no fine-tuning I've done so far has achieved that.