WeClone: Create an AI Avatar from Chat History

Overview

WeClone is a Python toolkit for building a conversational digital avatar from chat records. It provides an end-to-end workflow, from preparing message exports and fine-tuning a language model to serving the result through a chatbot.

It is aimed at individuals and developers who want more than a generic chatbot: a model adapted to a particular person's conversational style. The project offers local processing and deployment options, but users need to manage the data, model, hardware, and consent considerations themselves.

Key Features

• Preprocess chat history exported from Telegram, including text and supported image data.
• Fine-tune compatible language models with supervised fine-tuning and LoRA-based workflows.
• Filter selected personal information with Microsoft Presidio and a configurable blocklist.
• Optionally encrypt stored data and outputs, with password-based access.
• Build a user profile from processed conversations.
• Serve inference through an API and connect it to supported chatbot platforms.
• Provide a browser chat demo for testing model responses.

Use Cases

• Individuals can experiment with a personal conversational avatar based on their own exported chats.
• ML practitioners can study how fine-tuning on conversation data affects a model's response style.
• Bot developers can connect a custom model to messaging platforms through WeClone's inference service.
• Researchers can test privacy filtering and local model workflows on authorized conversation data.

Getting Started

The README recommends Python 3.12, CUDA 12.6 or above for CUDA setups, and uv for installing dependencies:

git clone https://github.com/xming521/WeClone.git && cd WeClone
uv venv .venv --python=3.12
source .venv/bin/activate
uv pip install --group main -e .

See the README and project documentation for configuration, data preparation, training, and deployment instructions.

Considerations

• Fine-tuning is resource-intensive. The README lists 16 GB VRAM for a 7B model with LoRA and 6 GB with 4-bit QLoRA; actual needs depend on configuration.
• The README says performance depends on the size and quality of the model and chat data, and describes 7B results as average. The project is still in rapid iteration.
• Telegram is the documented chat-history source. Other listed sources, including WhatsApp, Discord, and Slack, are marked as in progress for data import.
• Privacy filters are not guaranteed to catch every sensitive detail. Protect source chats, obtain appropriate authorization, and review outputs before use.
• The project is licensed under AGPL-3.0. Review the license and the project's usage disclaimer before distributing or operating a service based on it.