🚀 One-stop solution for creating your digital avatar from chat history 💡

简体中文｜ English｜ Project Homepage ｜ Documentation

Important

WhatsApp and Telegram chat logs integration for digital avatar creation is coming !

✨Core Features

• 💫 Complete end-to-end solution for creating digital avatars, including chat data export, preprocessing, model training, and deployment
• 💬 Fine-tune LLM using chat history with support for image modal data, infusing it with that authentic "flavor"
• 🔗 Integrate with Telegram, WeChat, WhatsApp (coming soon) to create your own digital avatar
• 🛡️ Privacy information filtering with localized fine-tuning and deployment for secure and controllable data

📋Features & Notes

Important

WeClone is currently not partnered with any platform and has not issued any cryptocurrency. The only official website is: weclone.love. Beware of imitations.

Chat Platform Support

Platform	Text	Images	Voice	Video	Animated Emojis	Links (Sharing)	Quote	Forward	Location	Files
WeChat	✅	✅	❌	❌	❌	❌	❌	❌	❌	❌
Telegram	✅	✅	❌	❌	⚠️Convert to Emoji	❌	❌	✅	✅	❌

Important

• WeClone is still in rapid iteration phase, current performance does not represent final results.
• LLM fine-tuning effectiveness largely depends on model size, quantity and quality of chat data. Theoretically, larger models with more data yield better results.
• 7B models are prone to becoming "dumb", 14B models can barely communicate, while 32B+ models perform much better.
• Windows environment has not been rigorously tested. You can use WSL as the runtime environment.

Recent Updates

[25/06/05] Support for image modal data fine-tuning

Hardware Requirements

The project uses Qwen2.5-VL-7B-Instruct model by default with LoRA method for SFT stage fine-tuning. You can also use other models and methods supported by LLaMA Factory.

Estimated VRAM requirements (text-only large model memory usage as follows, vision models increase based on image quantity and size):

Method	Precision	7B	14B	30B	70B	xB
Full (bf16 or fp16)	32	120GB	240GB	600GB	1200GB	18xGB
Full (pure_bf16)	16	60GB	120GB	300GB	600GB	8xGB
Freeze/LoRA/GaLore/APOLLO/BAdam	16	16GB	32GB	64GB	160GB	2xGB
QLoRA	8	10GB	20GB	40GB	80GB	xGB
QLoRA	4	6GB	12GB	24GB	48GB	x/2GB
QLoRA	2	4GB	8GB	16GB	24GB	x/4GB

Environment Setup

1. CUDA installation (skip if already installed, requires version 12.6 or above)

2. It is recommended to use uv to install dependencies, which is a very fast Python environment manager. After installing uv, you can use the following commands to create a new Python environment and install dependencies.

git clone https://github.com/xming521/WeClone.git && cd WeClone
uv venv .venv --python=3.10
source .venv/bin/activate # windows .venv\Scripts\activate
uv pip install --group main -e . 

3. Copy the configuration file template and rename it to settings.jsonc, and make subsequent configuration changes in this file:

cp examples/tg.template.jsonc settings.jsonc

Note

Training and inference related configurations are unified in the file settings.jsonc

4. Use the following command to test whether the CUDA environment is correctly configured and can be recognized by PyTorch (not needed for Mac):

  python -c "import torch; print('CUDA Available:', torch.cuda.is_available());"

5. (Optional) Install FlashAttention to accelerate training and inference: uv pip install flash-attn --no-build-isolation.

Model Download

It is recommended to use Hugging Face to download models, or use the following command:

git lfs install
git clone https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct models/Qwen2.5-VL-7B-Instruct

Data Preparation

Please use Telegram Desktop to export chat records. Select Photos for message types and JSON for format. You can export multiple contacts (group chat records are not recommended), then place the exported ChatExport_* in the ./dataset/telegram directory, meaning put different people's chat record folders together in ./dataset/telegram.

Data Preprocessing

• The project by default removes phone numbers, ID numbers, and emails from the data. A blocked words dictionary blocked_words is also provided in settings.jsonc, where you can add words or phrases that need to be filtered (the entire sentence containing the blocked words will be removed).

Important

🚨 Please be sure to protect personal privacy and do not leak personal information!

• Execute the following command to process the data. You can modify the make_dataset_args in settings.jsonc according to your own chat style.

weclone-cli make-dataset

• Currently supports time window strategy. Messages from a single person are combined into one sentence by commas based on single_combine_time_window, and Q&A pairs are matched based on qa_match_time_window.
• For training multimodal large models: Enable by adding images to include_type, and control image quantity and size through image_max_pixels and max_image_num parameters to reduce VRAM usage.
• For Image to Text: Add images to include_type and configure vision_api parameters. The system will use external multimodal models to convert images to text, and the final generated dataset is still used for training text-only LLM.
• You can enable the enable_clean option in clean_dataset to clean the data for better results (multimodal data is not currently supported). The current system supports using llm judge to score chat records, providing vllm offline inference and API online inference methods. You can enable API online inference mode by changing "online_llm_clear": false to true , and configure the corresponding base_url, llm_api_key, model_name and other parameters. All models compatible with OpenAI interface can be accessed.
• After obtaining the llm scoring score distribution, you can filter acceptable data by setting the accept_score parameter, and appropriately reduce the lora_dropout parameter in train_sft_args to improve the model's fitting effect.

配置参数并微调模型

• (可选)修改 settings.jsonc 的 model_name_or_path 和 template 选择本地下载好的其他模型。
• 修改per_device_train_batch_size以及gradient_accumulation_steps来调整显存占用。
• 可以根据自己数据集的数量和质量修改train_sft_args的num_train_epochs、lora_rank、lora_dropout等参数。

单卡训练

weclone-cli train-sft

多卡环境单卡训练，需要先执行 export CUDA_VISIBLE_DEVICES=0

多卡训练

取消settings.jsonc中deepspeed行代码注释，使用以下命令多卡训练：

uv pip install deepspeed
deepspeed --num_gpus=使用显卡数量 weclone/train/train_sft.py

使用浏览器demo简单推理

可以在这一步测试出合适的temperature、top_p值，修改settings.jsonc的infer_args后，供后续推理时使用。

weclone-cli webchat-demo

使用接口进行推理

weclone-cli server

使用常见聊天问题测试

不包含询问个人信息的问题，仅有日常聊天。测试结果在test_result-my.txt。

weclone-cli server
weclone-cli test-model

🖼️ 微调效果

Tip

QQ群内有部署好的Qwen2.5VL 32B Bot，可以体验效果。

使用Qwen2.5-14B-Instruct模型，大概3万条处理后的有效数据，loss降到了3.5左右的效果：

截图

🤖 部署到聊天机器人

LangBot

LangBot 是一个开源的接入全球多种即时通信平台的 LLM 机器人平台，支持Discord、Telegram、Slack等平台，适合各种场景使用。

1. 部署 LangBot
2. 在 LangBot 中添加一个机器人
3. 在模型页添加新模型，名称gpt-3.5-turbo，供应商选择 OpenAI，填写 请求 URL 为 WeClone 的地址，详细连接方式可以参考文档，API Key 任意填写。

6. 在流水线配置中选择刚才添加的模型，或修改提示词配置

AstrBot

AstrBot 是易上手的多平台 LLM 聊天机器人及开发框架 ✨ 平台支持 QQ、Telegram、微信、企微、飞书。

使用步骤：

1. 部署 AstrBot
2. 在 AstrBot 中部署消息平台
3. 执行 weclone-cli server 启动api服务
4. 在 AstrBot 中新增服务提供商，类型选择OpenAI，API Base URL 根据AstrBot部署方式填写（例如docker部署可能为http://172.17.0.1:8005/v1） ，模型填写gpt-3.5-turbo,API Key随意填写一个
5. 微调后不支持工具调用，请先关掉默认的工具，消息平台发送指令： /tool off all，否则会没有微调后的效果。
6. 根据微调时使用的default_system，在 AstrBot 中设置系统提示词。

Important

检查api_service的日志，尽量保证大模型服务请求的参数和微调时一致，tool插件能力都关掉。

7. 调整采样参数，例如temperature、top_p、top_k等 配置自定义的模型参数

📌 路线图

• 更丰富的上下文：包括上下文对话、聊天对象信息、时间等
• Memory 支持
• 支持多模态:已支持图片
• 数据增强
• 支持GUI
• 支持COT思考

问题解决

官方文档FAQ

同时建议使用DeepWiki解决问题。

❤️ 贡献代码

欢迎任何 Issues/Pull Requests！

你可以通过查看Issues或帮助审核 PR（拉取请求）来贡献。对于新功能的添加，请先通过 Issue 讨论。
开发环境：

uv pip install --group dev -e .
pre-commit install

项目使用pytest测试，pyright检查类型，ruff检查代码格式。

🙏 致谢

感谢以下代码贡献者和社区里其他成员的贡献

同时本项目受益于PyWxDump、LLaMA-Factory、AstrBot、LangBot等优秀开源项目。

⚠️ 免责声明

Caution

请勿用于非法用途，否则后果自负。

1. 使用目的

• 本项目仅供学习交流使用，请勿用于非法用途，请勿用于非法用途，请勿用于非法用途，否则后果自负。
• 用户理解并同意，任何违反法律法规、侵犯他人合法权益的行为，均与本项目及其开发者无关，后果由用户自行承担。

2. 使用期限

• 您应该在下载保存使用本项目的24小时内，删除本项目的源代码和程序；超出此期限的任何使用行为，一概与本项目及其开发者无关。

3. 操作规范

• 本项目仅允许在授权情况下使用数据训练，严禁用于非法目的，否则自行承担所有相关责任；用户如因违反此规定而引发的任何法律责任，将由用户自行承担，与本项目及其开发者无关。
• 严禁用于窃取他人隐私，严禁用于窃取他人隐私，严禁用于窃取他人隐私，否则自行承担所有相关责任。

4. 免责声明接受

• 下载、保存、进一步浏览源代码或者下载安装、编译使用本程序，表示你同意本警告，并承诺遵守它;

5. 禁止用于非法测试或渗透

• 禁止利用本项目的相关技术从事非法测试或渗透，禁止利用本项目的相关代码或相关技术从事任何非法工作，如因此产生的一切不良后果与本项目及其开发者无关。
• 任何因此产生的不良后果，包括但不限于数据泄露、系统瘫痪、侵犯隐私等，均与本项目及其开发者无关，责任由用户自行承担。

6. 免责声明修改

• 本免责声明可能根据项目运行情况和法律法规的变化进行修改和调整。用户应定期查阅本页面以获取最新版本的免责声明，使用本项目时应遵守最新版本的免责声明。

7. 其他

• 除本免责声明规定外，用户在使用本项目过程中应遵守相关的法律法规和道德规范。对于因用户违反相关规定而引发的任何纠纷或损失，本项目及其开发者不承担任何责任。

• 请用户慎重阅读并理解本免责声明的所有内容，确保在使用本项目时严格遵守相关规定。

请用户慎重阅读并理解本免责声明的所有内容，确保在使用本项目时严格遵守相关规定。

⭐ Star History

Tip

如果本项目对您有帮助，或者您关注本项目的未来发展，请给项目 Star，谢谢

克隆我们，保留灵魂的芬芳