iClone

Clone your friends by finetuning Mistral-7B on your iMessage chat history.

• Built for Apple Silicon MacBooks
• Uses Apple's MLX framework for GPU acceleration
• All data is fetched from and stays on your machine

Todo

• Create a basic CLI that lets you choose a chat to finetune on
• Group consecutive messages to create better training data structure
  • Include system prompt, ~5 messages, and prediction of the next message as a single training example
• Trim down the Apple starter code to only what's necessary

Setup

This copies your iMessage database to the db directory to avoid operating on the original file. This file might be kind of big, mine is ~1.5GB.

cp ~/Library/Messages/chat.db chat.db

Install dependencies:

pip install -r requirements.txt

Download Mistral-7B weights:

curl -o models/mistral-7B.tar https://files.mistral-7b-v0-1.mistral.ai/mistral-7B-v0.1.tar
tar -xf models/mistral-7B.tar
rm models/mistral-7B.tar

Convert the model to MLX format:

python convert.py \
    --torch-model models/mistral-7B/ \
    --mlx-model models/mlx-mistral-7B/

Generate some messages:

python lora.py --model <path_to_model> \
               --adapter-file <path_to_adapters.npz> \
               --num-tokens 50 \
               --prompt "table: 1-10015132-16
columns: Player, No., Nationality, Position, Years in Toronto, School/Club Team
Q: What is terrence ross' nationality
A: "

References

• https://github.com/ml-explore/mlx-examples/tree/main/lora
• https://twitter.com/iamgingertrash/status/1628495957632614400
• https://github.com/1rgs/MeGPT
• https://spin.atomicobject.com/search-imessage-sql
• https://www.izzy.co/blogs/robo-boys.html
• https://edwarddonner.com/2024/01/02/fine-tuning-an-llm-on-240k-text-messages/
• https://github.com/gavi/mlx-whatsapp