Skip to content

Latest commit

 

History

History
295 lines (202 loc) · 11.9 KB

Readme_en.md

File metadata and controls

295 lines (202 loc) · 11.9 KB
Raw

Stars Model HuggingFace Python

简体中文 | 繁體中文 | English | 日本語

Important

As of 2024.12.04, due to a configuration format update, please reset your configuration file if you've fetched updates to this repository. We apologize for any inconvenience caused.

Warning

The chatbot is trained in Simplified Chinese, which may limit its ability to handle English input effectively. Let us know if you are interested in a version fine-tuned with a machine-translated dataset in English!

Tip

This page's content might not always reflect the latest updates. Visit the Simplified Chinese page for the most recent information.

Introduction ✨

Muice is an AI chatbot, who actively engages in conversation and is trained using over 3,000 dialogue samples. The chatbot features a conversational style resembling a cheerful anime girl who enjoys sharing daily life anecdotes and greets users with unique messages every day. It is fine-tuned on ChatGLM2-6B and Qwen.

Features 🪄

✅ Automated Environment Setup: Easy, nearly hands-free installation process.

✅Fine-tuned Models Provided: Provide ChatGLM2-6B P-Uni V2 model and Qwen Qlora fine-tuning model fine tuned by myself with 3k+conversation data

✅ Scheduled and Random Conversations: Automatically initiates chats at specific times of the day.

✅ Command Support: Includes five commands to refresh, reply, and manage conversations.

✅ OFA Image Recognition: Can recognize and send memes or stickers.

✅ TTS Support: Uses fish-speech for voice synthesis (custom TTS models are under development).

✅ Multi-Language Support: Comprehensive multilingual documentation.

⬜ Optimized memory modules for long-term and short-term memory capabilities.

⬜ Improved logging mechanisms with automated bug report generation.

⬜ Comprehensive FAQ guide.

Getting Started 💻

Recommended Environment:

  • Python: Version 3.10+
  • Hardware: A GPU with at least 6GB VRAM (4GB VRAM for int4 quantization, 16GB RAM for CPU inference).

Automated Installation

Follow these steps for a quick setup:

  1. Download and unzip the latest code from Code -> Download ZIP.
  2. Run install_env.bat by double-clicking or using the following command:
.\install_env.bat

Note: The script sets up a Python virtual environment, no Conda required.

Manual Installation (Using Conda)

git clone https://github.com/Moemu/Muice-Chatbot
cd Muice-Chatbot
conda create --name Muice python=3.10.10 -y
conda activate Muice
pip install -r requirements.txt

For GPU users:

pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu124

Ensure CUDA is configured properly ( Reference ).

Model Download and Loading

Supported base models:

Model Fine-tuned Version Extra Dependencies
ChatGLM2-6B-Int4 2.2-2.4 cpm_kernels
ChatGLM2-6B 2.0-2.3
Qwen-7B-Chat-Int4 2.3、2.6.2 llmtuner
Qwen2-1.5B-Instruct-GPTQ-Int4 2.5.3 llmtuner
Qwen2.5-7B-Instruct-GPTQ-Int4 2.7.1 llmtuner
RWKV (By Seikaijyu) see HF （RWKV-Runner is required）

Place the models in the model folder. Ensure fine-tuned models include a .model file.

Loading methods:

  • Via API
  • Using transformers.AutoTokenizer and transformers.AutoModel.
  • Through llmtuner.chat.ChatModel.
  • With RWKV-Runner APIs.

In the tested models, we recommend using the corresponding loading method for the following models, and other models can also be loaded using similar methods:

Base model Fine tuning method Loading method
ChatGLM P-tuning V2 transformers
Qwen sft llmtuner
RWKV (Seikaijyu fine-tuning) pissa rwkv apin

The model loading method can be adjusted in the configuration file:

# model
model:
  loader: transformers # transformers/llmtuner/rwkv-api
  model_path: ./model/chatglm2-6b # base model path
  adapter_path: ./model/Muice # fine-tuning model path

(If loading for API / rwkv-api , fill in the corresponding API address for model_name_or_path)

If you do not have a suitable graphics card and need to load models or quantization models through the CPU, please install and configure the GCC environment and check openmp. Reference link(ZH)

Bot Service Configuration ⚙️

This project supports the OneBot V11 Protocol, enabling compatibility with QQ and other applications.

To connect with QQ:

  1. Install LLOneBot.
  2. Enable reverse WebSocket and set it to ws://127.0.0.1:21050/ws/api.
  3. Alternatively, use Lagrange.Core or other adapters compatible with OneBot V11 (Adapter List).

Important: Avoid upgrading QQNT unnecessarily. If issues arise, consider downgrading.

Additional Features 🎨

  • Voice Reply: Supports TTS replies. Learn more(ZH).
  • Image Recognition: Handles meme recognition and response. Details(ZH).
  • Faiss Long-Term Memory: Experimental long-term memory capability. Documentation(ZH).
  • Real-Time Voice Chat: Allows for interactive voice communication. Setup Guide(ZH).

Project Structure Overview 🗂️

After completing setup, your directory should resemble the following:

Muice-Chatbot     <- Root directory
 ├─llm
 ├─model
 │  ├─ chatglm2-6b       <- Base model (select one)
 │  ├─ chatglm2-6b-int4  <- Int4 quantized model
 │  ├─ Qwen-7B-Chat-Int4 <- Qwen-7B-int4
 │  └─ Muice
 ├─configs.yml  <- Configuration file
 ├─ws.py         <- WebSocket service
 ├─main.py       <- Main application
 ├─requirements.txt
 └─...

Configuration ⚒️

The configuration file is located in configs.yml. Adjust settings to match your requirements.

Updates (Post-2024.12.04):

New configurations have been introduced for models versioned 2.7.x and above:

# Active Conversations
active:
  enable: false  # Enable or disable active conversations
  rate: 0.003    # Probability of initiating a conversation (per minute)
  active_prompts:
    - '<生成推文: 胡思乱想>'
    - '<生成推文: AI生活>'
    - '<生成推文: AI思考>'
    - '<生成推文: 表达爱意>'
    - '<生成推文: 情感建议>'
  not_disturb: true  # Enable Do Not Disturb mode
  schedule:
    enable: true  # Enable scheduled tasks
    rate: 0.75    # Probability of executing scheduled tasks
    tasks:
      - hour: 8
        prompt: '<日常问候: 早上>'
      - hour: 12
        prompt: '<日常问候: 中午>'
      - hour: 18
        prompt: '<日常问候: 傍晚>'
      - hour: 22
        prompt: '<日常问候: 深夜>'
  targets:  # List of QQ IDs for active conversations
    - 12345678
    - 23456789

For pre-2.7.x models, use the following configuration format instead:

  active_prompts:
    - '（分享一下你的一些想法）'
    - '（创造一个新话题）'

And:

    tasks:
      - hour: 8
        prompt: '（发起一个早晨问候）'
      - hour: 12
        prompt: '（发起一个中午问候）'
      - hour: 18
        prompt: '（发起一个傍晚问候）'
      - hour: 22
        prompt: '（发起一个临睡问候）'

Usage 🎉

Run the main application from the project root directory:

conda activate Muice
python main.py

Alternatively, use the generated start.bat script.

Commands 🕹️

Command Description
/clean Clear the current conversation history
/refresh Refresh the current conversation
/help Display all available commands
/reset Reset all conversation data
/undo Undo the last conversation entry

Example Conversations (Training Data) 📑

View the public dataset: Moemu/Muice-Dataset.

Character Design 🎨

Unlike other chatbot projects, this project provides a model that I have fine tuned based on my own conversation dataset, which can be downloaded in the Release. The publicly available information regarding the fine tuned model persona is as follows:

Muice Character Design

Training set open source address: Moemu/Muice-Dataset

Original model: THUDM/ChatGLM2-6B & QwenLM/Qwen

The source code of this project uses [MIT License]（ https://github.com/Moemu/Muice-Chatbot/blob/main/LICENSE ）For fine tuned model files, it is not recommended to use them for commercial purposes.

About🎗️

**Friendship Link: **Coral

All contributors:

Loading...

If you find this project helpful, consider supporting it:

Buy Me A Coffee

Thanks you all!

Star History:

Star History Chart