I fine-tuned a 2B LLM on our WhatsApp group chat, and shared how to do it on GitHub as a cookbook.

https://github.com/Sayitobar/chat_llm_cookbook
This is my personal project that took several months. I wanted to see whether a 2B small local model could simulate a six-person group chat trained & ran on an M1 Pro.

How good is it?:
- It's fun, but not great. It doesn't achieve coherent & consistent group simulation, but it learned enough of our slang, reactions, and pacing to be fun. The generated messages are very similar to what we'd type.
- There is some coherence, but not a deep understanding, and the model doesn't hold information about us (expect our names and a few very obvious stuff).

How good is it on paper?:
- I have evaluated my models performances by judging them with a judge LLM. The best version achieved an 80% human win rate at human-vs-model tests, ideal should be <50%.

The main thing is, you'll have a lot of fun chatting with this model once you train it on YOUR data. (ask for consent pls)

Cookbook:
I published the reproducible local pipeline, chat UI, human-anchored evaluation, results, and experiment PDF. No private chat data or fine-tuned weights are released :)

Fyi, all of the tests I've done were in Turkish.

This project is still unfinished as there are still architectures and training data formats I haven't tested, or stronger 2B models that aren't released yet, as of September 2026.