ChatGLM: Bilingual Open Model

Overview ChatGLM is a family of open-source LLMs developed by Tsinghua University (KEG & Data Mining Group) in China. It is specifically optimized for Chinese-English bilingual conversation.

Architecture (GLM) It uses the GLM (General Language Model) architecture, which is different from GPT (Decoder-only) or BERT (Encoder-only). It uses Autoregressive Blank Infilling.

Key Models

• ChatGLM-6B: A 6 billion parameter model (very small!).
• Quantization: Famous for running on consumer hardware (consumer laptop GPUs). It can run in 4-bit mode on just 6GB VRAM.

Significance Before Llama 2 became the global standard, ChatGLM was the go-to open model for the Chinese NLP community. It handles Chinese idioms, culture, and grammar significantly better than western-trained models (GPT-3).

Deployment It is fully integrated into Hugging Face Transformers.

model = AutoModel.from_pretrained("THUDM/chatglm-6b", trust_remote_code=True)
response, history = model.chat(tokenizer, "你好", history=[])