sorokinvld/megpt

Fine-tune a large language model on your own iMessages

MeGPT: Fine-tune a Language Model with your iMessage Conversations

MeGPT allows you to fine-tune a large language model on your own messages, enabling you to talk to yourself.

This repo contains code for:

1. Extracting your iMessage conversations from your Mac
2. Fine-tuning a large language model on your messages
3. Generating completions using the fine-tuned model

This is a sample repo that trains Meta AI's OPT 1.3b model with Parallel Efficient Fine-tuning (PEFT) on your iMessage conversations. You can use this repo as a starting point for fine-tuning other models on your own data.

Based off of example code from lvwerra/trl

Usage

1. Install the requirements:

pip install -r requirements.txt

1. On your Mac, run extract_messages.py to extract your iMessage conversations and save them to a CSV file:

python extract_messages.py

1. Configure fine_tune.py with your desired model, input CSV, and other settings. For example:

model_name = "facebook/opt-1.3b"
block_size = 128
input_csv = "messages.csv"

To see the full list of supported models, visit PEFT Models Support Matrix.

1. Train the model on your messages using fine_tune.py:

python fine_tune.py

1. Generate completions using the fine-tuned model with generate.py:

python generate.py