[luke-stanley]

For my ChillTranslator project I spent maybe a few dollars fine-tuning Phi 2, to generate less spicy variations of inflammatory Hacker News comments with very little data to see how well it worked (especially compared to your 100M tokens). I'll improve it when I have time. I mostly followed the Brev fine-tune tutorial but I wanted to have a 2 GB file GGUF quantised model I could run on any device with a specific JSON grammar. It uses Transformers PEFT and QLoRA. I didn't try Axolotl yet, or OpenPipe but I hope to. Actual compute time is probably much less than I spent, I wasted time dealing with drivers, trying to figure out how to merge the finetuned weights, serialise to old fashioned Pickle, not safe-tensors, and how to convert to GGUF, quantise it and rsync it.