WeClone is a one-stop open-source solution for creating an AI digital avatar from chat records, hosted on GitHub (https://github.com/xming521/WeClone）。其核心思路是：导出聊天数据，经过预处理后微调大语言模型，使模型捕捉个人独特的说话风格，再绑定到聊天机器人，从而让数字自我“活”起来。
In terms of use, WeClone is aimed at users who want to build a personalized AI digital avatar based on their own chat history. It provides a complete end-to-end process from chat data export, preprocessing, and model training to deployment, and supports fine-tuning with image modality data, allowing image information in chat records to be included in training as well.
Main capabilities include: fine-tuning an LLM using chat history, with support for image modality data; Telegram integration is supported, and WhatsApp integration is coming soon; privacy information filtering is provided, and localized fine-tuning and deployment are supported to ensure data security and controllability. The project uses the Qwen2.5-VL-7B-Instruct model by default and uses the LoRA method for SFT-stage fine-tuning, but other models and methods supported by LLaMA Factory can also be used. Data source platforms support Telegram (added on July 10, 2025), and deployment platforms support AstrBot, LangBot, and others.
For usage, the project recommends using uv to install dependencies, and requires CUDA 12.6 or above (not required on Mac). The configuration file template must be copied and renamed to settings.jsonc, and training and inference-related configurations are all unified in this file. For data preparation, Telegram Desktop must be used to export chat records, selecting the Photos message type and JSON format, and placing the exported ChatExport_* into the ./dataset/telegram directory. During the data preprocessing stage, configurations such as language, platform, and include_type can be modified, and Telegram users need to set telegram_args.my_id. The project uses Microsoft Presidio by default to remove phone numbers, email addresses, credit card numbers, IP addresses, geographic location names, international bank account numbers, cryptocurrency wallet addresses, age information, and generic ID numbers, but cannot guarantee 100% recognition, and also provides a blocked_words blacklist for manual filtering. Data encryption and passwords are set through security_args. New templates are encrypted by default, encrypted output ends with .enc, and the .weclone/ directory must be retained to read encrypted data. Training can be single-GPU or multi-GPU (the deepspeed line in settings.jsonc must be uncommented), and inference can be done through a browser demo or API. When deploying to a chatbot, you can run weclone-cli server --inference --port 8005 to start the API service, and add an OpenAI-type provider in AstrBot or LangBot, filling in gpt-3.5-turbo for the model and any value for the API Key. Note that tool calling is not supported after fine-tuning, so the default tools must be disabled first.
Applicable scenarios include: personal digital avatar experiments, chat style simulation, multi-platform chatbot integration, and more. In terms of limitations: the project is still in a stage of rapid iteration, and current performance does not represent the final result; the fine-tuning effect depends largely on model size and the quantity and quality of chat data, and in theory, the larger the model and the more data, the better the effect; the 7B model performs average, while models with 14B parameters and above usually perform better; the Windows environment has not been strictly tested, and WSL can be used as the runtime environment. The project is for learning, research, and experimental purposes only, and use in a production environment carries significant risks that must be assessed by the user. When using a digital avatar, it is strongly recommended to clearly label it as an "AI Bot" or "digital avatar," and to prominently mark "AI-generated content" in the interface to avoid users mistaking it for a conversation with a real person. The project is provided "as is," without any express or implied warranties, and the author does not promise ongoing technical support or maintenance.
