CodeCapybara: Code Writing LLaMa Finetuned on Deepmind Dataset

Its LLaMA 7B + Deepmind CodeContests + Alpaca, and they are attempting to replicate Facebook's evaluations. Looks like the repo is a WIP.

There is a 4-bit Triton quantization on HF, which I am playing with on a 6GB GPU right now: https://huggingface.co/catalpa/codecapybara-4bit-128g-gptq

You might want to format queries like Alpaca + the tests in the deepmind dataset: https://huggingface.co/datasets/deepmind/code_contests/tree/...