~120,000 code instruction/response pairs in Alpaca format were trained on the base model to achieve this result.

Very curious where they are getting this data from. In other open source papers, usually this comes from a GPT-4 output, but presumably Stability would not do that?