- Hosted application: https://dollyai.app
- Completely rewritten with Torchtune library
- ~30% reduced vRAM usage
- Much simpler codebase
- Llama3 support
This repository lets you create an AI chatbot clone of yourself, using your WhatsApp chats as training data. It builds upon the new torchtune library for finetuning and inference.
It includes code to:
- Preprocess exported WhatsApp chats into a suitable format for finetuning
- Finetune a model on your WhatsApp chats, using (quantized) LoRa
- Chat with your finetuned AI clone, via a commandline interface
Currently supported models are:
- Llama3-8B-Instruct
- Mistral-7B-Instruct-v0.2
- Clone this repository
- Ensure you have pytorch installed in your active environment. If not, follow these instructions: https://pytorch.org/get-started/locally/
- Install torchtune:
git clone https://github.com/kinggongzilla/torchtune.git
cd torchtune
pip install .
cd ..
Note that slight modifications to the torchtune library ChatDataset class code were necessary, hence we're not installing from the official repo. In particular the validate_messages function call is removed, to allow for message threads which are not strictly alternating between human and assistant roles.
Run tune download meta-llama/Meta-Llama-3-8B-Instruct --output-dir model/llama3 --hf-token <HF_TOKEN>
. Replace <HF_TOKEN> with your hugging face access token. In order to download Llama3 you first need to request access on the Meta Llama3 Huggingface page.
Run tune download mistralai/Mistral-7B-Instruct-v0.2 --output-dir model/mistral
.
If you downloaded the model in another format (e.g. safetensors), please adjust the checkpoint_files in mistral/qlora_train_config.yaml
.
To prepare your WhatsApp chats for training, follow these steps:
- Export your WhatsApp chats as .txt files. This can be done directly in the WhatsApp app on your phone, for each chat individually. You can export just one .txt from a single chat or many .txt files from all your chats.
Unfortunately, formatting seems to vary between regions. I am based on Europe, so the regex in the
preprocess.py
might have to be adjusted if you are based in a different region. - Copy the .txt files you exported into
data/raw_data
. - Run
python preprocess.py "YOUR NAME"
. This will convert your raw chats into a sharegpt format suitable for training and saves the JSON files todata/preprocessed
. ReplaceYOUR NAME
with the exact string which represents your name in the exportet WhatsApp .txt files. The script will assign you the "gpt" role and your conversation partners the "user" role.
Run tune run lora_finetune_single_device --config config/llama3/qlora_train_config.yaml
Run tune run lora_finetune_single_device --config config/mistral/qlora_train_config.yaml
Run tune run chat.py --config config/llama3/inference_config.yaml
You can define your own system prompt by changing the prompt
string in the config/llama3/inference_config.py
file.
For mistral to fit onto 24GB I first had to quantize the trained model.
- Run
tune run quantize --config config/mistral/quantization.yaml
- Run
tune run chat.py --config config/mistral/inference_config.yaml
Running this command loads the finetuned model and let's you have a conversation with it in the commandline.
Approx 16 GB vRAM required for QLoRa Llama3 finetune with 4k context length. I ran the finetune on a RTX 3090. When experimenting with other models, vRAM requirement might vary.
- I trained my model but it did not learn my writing style
Try training for more than one epoch. You can change this in the
qlora_train_config.yaml
file. - The preprocessing script does not work
You probably need to adjust the regex pattern in
preprocess.py
. The WhatsApp export format varies from region to region. - I want to train a clone of on group chats
The current setup does not support group chats. Hence do not export and save them into the
data/raw_data
directory. If you do want the model to simulate group chats, I think you have to adjust the preprocessing and ChatDataset of torchtune, such that they support more than 2 roles. I haven't tried this myself.
- After training, adjust temperature and top_k parameters in the
inference_config.yaml
file. I found a temperature of 0.2 and top_k of 10000 to work well for me. - Finetuning works best with English chats. If your chats are in another language, you may need to adjust the preprocessing and training parameters accordingly.