This is a weekend experiment in which I finetuned a 7B 4-bit quantized LLM on my Discord messages.
The end result was a fairly convincing and pretty hilarious toy that I shared with a few friends.
I published a blog post with what I learned here.
./discord-bot
: a small discord bot wrapper in Go./model-training
: a set of Python scripts / notebooks for data prep and finetuning the QLoRA adapter
none :)
unfortunately I don't have the bandwidth for feature requests or support questions! I will, however, accept PRs of any kind, which will be reviewed on a "best-effort" basis.