ChatGLM3

ChatGLM3 Project Introduction

Overview

ChatGLM3 is an innovative conversational pre-training model co-developed by Zhipu AI and the Knowledge Engineering Group (KEG) at Tsinghua University. As an open-source initiative, ChatGLM3-6B represents the latest in a series of models aiming to advance the field of conversational AI by offering enhanced features while maintaining the accessible deployment and fluid dialogue capabilities of its predecessors.

Core Features of ChatGLM3-6B

1. Enhanced Base Model
   The foundation of ChatGLM3-6B, known as ChatGLM3-6B-Base, incorporates a more diverse set of training data, an increased number of training iterations, and improved training strategies. Evaluations across datasets involving semantics, mathematics, reasoning, coding, and knowledge demonstrate its leading performance among models with parameters under 10 billion.

2. Comprehensive Functional Support
   ChatGLM3-6B introduces a newly designed prompt format that extends beyond typical multi-turn conversations. It natively supports tool usage (known as Function Call), code execution, and complex scenarios involving agent tasks.

3. Expanded Open-source Models
   Beyond ChatGLM3-6B, the series includes the ChatGLM3-6B-Base model, the long-text conversation model ChatGLM3-6B-32K, and a model further enhanced for long-text understanding, ChatGLM3-6B-128K. These models are fully open for academic research and available for free commercial use upon registration.

Contributions and Responsibilities

The release of the ChatGLM3 open-source models aims to foster advancements in large model technologies within the open-source community. Developers and users are encouraged to adhere to the stipulated open-source agreements, ensuring these models and their derivatives are not used harmfully or in unmanaged services. Notably, the project team itself has not developed any commercial applications based on the ChatGLM3 models.

Model Availability

ChatGLM3 models can be accessed and downloaded from various platforms:

• ChatGLM3-6B: HuggingFace, ModelScope, WiseModel, OpenXLab
• Additional models like ChatGLM3-6B-Base, ChatGLM3-6B-32K, and ChatGLM3-6B-128K are also available on these platforms with varying sequence lengths catering to different user needs.

Community Engagement and Support

The project has gained support from various outstanding open-source platforms focused on inference acceleration and efficient fine-tuning, including:

• Inference Acceleration: Projects like chatglm.cpp, ChatGLM3-TPU, TensorRT-LLM by NVIDIA, and Intel’s OpenVINO offer pathways for deploying ChatGLM3-6B with enhanced speed and efficiency on different devices.
• Efficient Fine-Tuning: Initiatives such as LLaMA-Factory provide frameworks for easily refining the models.

Application Frameworks

Several application frameworks enhance the capabilities of ChatGLM3:

• LangChain-Chatchat: An open-source project combining ChatGLM with Langchain to create offline deployable, retrieval-augmented generation (RAG) knowledge bases.
• BISHENG: A platform dedicated to accelerating the practical application of large models.
• RAGFlow: An engine for retrieval-augmented generation based on deep document understanding for reliable question-answering.

Evaluation and Performance

In terms of performance, the ChatGLM3-6B version outperforms its predecessors and comparably sized models across several benchmarks, demonstrating significant improvements in mathematical reasoning, general knowledge, and long-text comprehension tasks.

Getting Started

To experiment with ChatGLM3, users can clone the repository, install dependencies, and explore various demonstration modes like chat interfaces, tool utilization examples, and code execution environments. Detailed guidelines for downloading and leveraging these capabilities are provided within the project’s documentation.

Through such robust features and extensive support, ChatGLM3 models aim to significantly contribute to the advancement of conversational AI and large language models. By inviting community collaboration and ensuring ample resources, the project seeks to become a cornerstone in the open-source AI landscape.