简体中文｜ English｜ Project Homepage ｜ Documentation
> [!IMPORTANT] > ### Telegram is now supported as a data source ! ## ✨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, 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 ### Data Source Platform Support | Platform | Text | Images | Voice | Video | Animated Emojis/Stickers | Links (Sharing) | Quote | Forward | Location | Files | |----------|------|--------|-------|-------|-----------------|-----------------|-------|---------|----------|-------| | Telegram | ✅ | ✅ | ❌ | ❌ | ⚠️Convert to Emoji | ❌ | ❌ | ✅ | ✅ | ❌ | | WhatsApp | 🚧 | 🚧 | 🚧 | 🚧 | 🚧 | 🚧 | 🚧 | 🚧 | 🚧 | 🚧 | | Discord | 🚧 | 🚧 | 🚧 | 🚧 | 🚧 | 🚧 | 🚧 | 🚧 | 🚧 | 🚧 | | Slack | 🚧 | 🚧 | 🚧 | 🚧 | 🚧 | 🚧 | 🚧 | 🚧 | 🚧 | 🚧 | ### Deployment Platform Support | Platform | Deployment Support | |----------|--------------------| | Telegram | ✅ | | WhatsApp | 🚧 | | Discord | ✅ | | Slack | ✅ | > [!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. > - The performance of the 7B model is average, while models with 14B or more parameters tend to deliver better results. > - Windows environment has not been rigorously tested. You can use WSL as the runtime environment. ### Recent Updates [25/07/10] Data source added Telegram [25/06/05] Support for image modal data fine-tuning ### Online Fine-Tuning - Big Model Lab (Lab4AI) (with 50 CNY voucher): https://www.lab4ai.cn/project/detail?utm_source=weclone1&id=ab83d14684fa45d197f67eddb3d8316c&type=project ### 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](https://github.com/hiyouga/LLaMA-Factory/tree/main#supported-models). Estimated VRAM requirements: | Method | Precision | 7B | 14B | 30B | 70B | `x`B | | ------------------------------- | --------- | ----- | ----- | ----- | ------ | ------- | | Full (`bf16` or `fp16`) | 32 | 120GB | 240GB | 600GB | 1200GB | `18x`GB | | Full (`pure_bf16`) | 16 | 60GB | 120GB | 300GB | 600GB | `8x`GB | | Freeze/LoRA/GaLore/APOLLO/BAdam | 16 | 16GB | 32GB | 64GB | 160GB | `2x`GB | | QLoRA | 8 | 10GB | 20GB | 40GB | 80GB | `x`GB | | QLoRA | 4 | 6GB | 12GB | 24GB | 48GB | `x/2`GB | | QLoRA | 2 | 4GB | 8GB | 16GB | 24GB | `x/4`GB | ## Environment Setup 1. CUDA installation (skip if already installed, **requires version 12.6 or above**) 2. It is recommended to use [uv](https://docs.astral.sh/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. ```bash 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: ```bash 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): ```bash 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](https://huggingface.co/docs/hub/models-downloading) to download models, or use the following command: ```bash 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](https://desktop.telegram.org/) to export chat records. Click the top right corner in the chat interface, then click "Export chat history". 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 - First, modify the `language`, `platform`, and `include_type` in the configuration file according to your needs. - If you use telegram, you need to modify the `telegram_args.my_id` in the configuration file to your own telegram user ID. - By default, the project uses Microsoft Presidio 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` from the data, but it cannot guarantee 100% identification. - Therefore, a blocklist `blocked_words` is provided in `settings.jsonc`, allowing users to manually add words or phrases they want to filter (the entire sentence containing blocked words will be removed by default). > [!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. ```bash weclone-cli make-dataset ``` More Parameter Details: [Data Preprocessing](https://docs.weclone.love/docs/deploy/data_preprocessing.html#related-parameters) ## Configure Parameters and Fine-tune Model - (Optional) Modify `model_name_or_path`, `template`, `lora_target` in `settings.jsonc` to select other locally downloaded models. - Modify `per_device_train_batch_size` and `gradient_accumulation_steps` to adjust VRAM usage. - You can modify parameters like `num_train_epochs`, `lora_rank`, `lora_dropout` in `train_sft_args` based on your dataset's quantity and quality. ### Single GPU Training ```bash weclone-cli train-sft ``` ### Multi-GPU Training Uncomment the `deepspeed` line in `settings.jsonc` and use the following command for multi-GPU training: ```bash uv pip install "deepspeed<=0.16.9" deepspeed --num_gpus=number_of_gpus weclone/train/train_sft.py ``` ### Simple Inference with Browser Demo Test suitable temperature and top_p values, then modify `infer_args` in settings.jsonc for subsequent inference use. ```bash weclone-cli webchat-demo ``` ### Inference Using API ```bash weclone-cli server ``` ### Test with Common Chat Questions Does not include questions asking for personal information, only daily conversation. Test results are in test_result-my.txt. ```bash weclone-cli server weclone-cli test-model ``` ## 🖼️ Results Showcase > [!TIP] > **We're looking for interesting examples of native English speakers chatting with WeClone! Feel free to share them with us on Twitter.** ## 🤖 Deploy to Chat Bots ### AstrBot [AstrBot](https://github.com/AstrBotDevs/AstrBot) is an easy-to-use multi-platform LLM chatbot and development framework ✨ Supports Discord, Telegram, Slack, Feishu and other platforms. Usage steps: 1. Deploy AstrBot 2. Deploy messaging platforms like Discord, Telegram, Slack in AstrBot 3. Execute `weclone-cli server` to start the API service 4. Add a new service provider in AstrBot, select OpenAI type, fill in the API Base URL according to AstrBot's deployment method (e.g., for docker deployment it might be http://172.17.0.1:8005/v1), fill in the model as gpt-3.5-turbo, and enter any API Key 5. Tool calling is not supported after fine-tuning, please turn off the default tools first by sending the command: `/tool off_all` on the messaging platform, otherwise the fine-tuned effect won't be visible. 6. Set the system prompt in AstrBot according to the default_system used during fine-tuning.  > [!IMPORTANT] > Check the api_service logs to ensure that the large model service request parameters are consistent with those used during fine-tuning as much as possible, and turn off all tool plugin capabilities. ### LangBot [LangBot](https://github.com/langbot-app/LangBot) is an easy-to-use open-source LLM chatbot platform suitable for various scenarios. It connects to various global instant messaging platforms. You can set up your IM bot in just 5 minutes.