ChatPLUG is an open-domain dialogue system designed for Chinese language users, offering personalized conversational experiences and multi-task capabilities. It targets researchers and developers looking to build custom chatbots or integrate advanced dialogue features into applications. The system's key benefits include knowledge augmentation, personalization through role-playing, and multi-skill proficiency.

How It Works

ChatPLUG leverages a large language model architecture, enhanced with techniques for knowledge augmentation and personalization. Knowledge augmentation allows for real-time integration of external information, such as search engine results or local knowledge bases, to provide up-to-date and domain-specific responses. Personalization is achieved through customizable bot profiles and role-play instructions, enabling users to define character traits and conversational styles.

Quick Start & Requirements

• ModelScope: Download checkpoints and use CLI for inference.
• HuggingFace: Coming soon for inference.
• XDPX Library:
  • Install: pip install -e . within the XDPX directory.
  • Requirements: CUDA, Python. bf16/fp16 support recommended for training.
  • Inference: CUDA_VISIBLE_DEVICES=0 x-script fidchat_new chat_pipeline/chatplug_3.7B_sftv2.6.0_instruction.hjson input
  • Training: Requires dataset preprocessing and specific DeepSpeed configurations.
  • Docs: User Guides

Highlighted Details

• Offers models like ChatPLUG-240M and ChatPLUG-3.7B on ModelScope.
• XDPX library provides streamlined tools for custom model training and chatbot building.
• Supports retrieval and role-play functionalities.
• Future releases planned for ModelScope and HuggingFace.

Maintenance & Community

The project is actively developed, with recent updates including training code and guides for retrieval and role-play. Links to ModelScope for demos and checkpoints are provided.

Licensing & Compatibility

Licensed under the Apache License (Version 2.0), permitting commercial use and integration with closed-source projects.

Limitations & Caveats

Training and deployment functionalities are marked as "Not Ready" for ModelScope and HuggingFace integrations, with XDPX being the primary method for these advanced features. Deployment options are listed as "Coming soon."