# How an LLM‘s fine-tuning data looks like

Source: https://twitter.com/levelsio/status/1727104127988044236

## Summary

The poster explains that ChatGPT-style chat models are instruct models, analogous to instruct-pix2pix in Stable Diffusion, which edits images from text commands like "change the haircolor to red," but applied to text. They describe fine-tuning as much simpler than expected: a JSON file of question-and-answer pairs that teaches a language model to reply to questions with answers. The poster notes this is nearly the same technique used for Dreambooth or LoRA fine-tuning on photos of people, except applied to text rather than images. They link to the Stanford Alpaca repository, which fine-tunes LLaMA, a model they describe as Facebook's leaked GPT model.

## Article

With all the OpenAI drama, time for me to try self hosting Alpaca, a finetuning ChatGPT-type model of LLaMA, which is Facebook's leaked GPT model

Now I finally get the origins of ChatGPT, it's just instruct models, same as instruct-pix2pix we had in Stable Diffusion, e.g. "change the haircolor to red" but for text

And instruct-text is what OpenAI smartly branded as "ChatGPT" and that became huge

Fun to see the technical origins of something so big were just available to us all

The finetuning is way way way way easier than I thought, not some magic but just a JSON text file with questions and answers which teachers the LLM model to reply to questions with answers, e.g. a chat

It's exactly the same what I use for Photo AI etc. finetuning with Dreambooth or LoRa, it's almost exactly the same tech, but you finetune on photos of people instead of text files, obviously I knew that but interesting to actually see it in front of you

https://github.com/tatsu-lab/stanford_alpaca