ChatGLM, the Tsinghua-based Gigabit base conversation model, has launched internal testing and is an open source single-card version of the model.

The release of ChatGPT has stirred up the entire AI field, and major technology companies, startups, and university teams are following suit. Recently, Heart of the Machine has reported on the research results of many startup companies and university teams.

Yesterday, another large-scale domestic AI dialogue model made its grand debut: ChatGLM, a company’s smart spectrum AI transformed from Tsinghua’s technological achievements and based on the GLM-130B 100 billion base model, is now open for invitations Internal testing.

It is worth mentioning that Zhipu AI has also open sourced the Chinese-English bilingual dialogue model ChatGLM-6B, which supports inference on a single consumer-grade graphics card.

#Internal beta application website: chatglm.cn

It is understood that the capability improvement of the current version of ChatGLM model mainly comes from the unique 100 billion base model GLM-130B. It is an architecture different from BERT, GPT-3 and T5. It is an autoregressive pre-training model containing multi-objective functions.

In August 2022, Tsinghua University and Zhipu AI opened the Chinese-English bilingual dense model GLM-130B with 130 billion parameters to the research community and industry. This model has some unique features Advantages:

• Bilingual: Supports both Chinese and English;
• High precision (English): in the public English natural language Better than GPT-3 175B (API: davinci, base model), OPT-175B and BLOOM-176B on the list of LAMBADA, MMLU and Big-bench-lite;
• High accuracy (Chinese): Significantly better than ERNIE TITAN 3.0 260B and YUAN 1.0-245B on 7 zero-sample CLUE data sets and 5 zero-sample FewCLUE data sets;
• Fast inference: First A 100 billion model that implements INT4 quantification, supporting fast and basically lossless inference with a 4-card 3090 or 8-card 2080Ti server;
• Reproducibility: all results (more than 30 Each task) can be reproduced through our open source code and model parameters;
• Cross-platform: supports domestic Haiguang DCU, Huawei Ascend 910 and Sunway processors and the United States Training and inference on NVIDIA chips.

Now, referring to the design ideas of ChatGPT, ChatGLM has injected code pre-training into the GLM-130B base model, through supervised fine-tuning (Supervised Fine-Tuning) and other technologies to achieve alignment of human intentions.

Heart of the Machine has obtained the internal beta invitation code. Here is a simple conversation with ChatGLM. The effect is as follows:

It can understand the actual meaning of "station CP":

Give ChatGLM a math problem to try:

Since I learned the linear equation of two variables, basic "chicken and rabbit in the same cage" problems like this can no longer trouble it:

Open Source ChatGLM-6B

ChatGLM-6B is an open source dialogue language model that supports bilingual question and answer in Chinese and English and is optimized for Chinese. The model is based on the General Language Model (GLM) architecture and has 6.2 billion parameters. Combined with model quantization technology, users can deploy it locally on consumer-grade graphics cards (a minimum of 6GB of video memory is required at the INT4 quantization level). ChatGLM-6B uses the same technology as ChatGLM and is optimized for Chinese Q&A and dialogue. After bilingual training in Chinese and English with about 1T identifiers, supplemented by supervised fine-tuning, feedback self-service, human feedback reinforcement learning and other technologies, the 6.2 billion parameter ChatGLM-6B, although not as large as the 100 billion model, has greatly reduced the inference cost and improved It has improved efficiency and can already generate answers that are quite consistent with human preferences.

Model open source address: https://github.com/THUDM/ChatGLM-6B

Specifically , ChatGLM-6B has the following characteristics:

• # Sufficient bilingual pre-training in Chinese and English: ChatGLM-6B has trained a token amount of 1T on Chinese and English materials in a 1:1 ratio. , both bilingual.
• Optimized model architecture and size: Drawing on GLM-130B training experience, the two-dimensional RoPE position encoding implementation is revised, using the traditional FFN structure. The parameter size of 6B (6.2 billion) also makes it possible for researchers and individual developers to fine-tune and deploy ChatGLM-6B themselves.
• Lower deployment threshold: At FP16 half precision, ChatGLM-6B requires at least 13 GB of video memory for inference. Combined with model quantization technology, this requirement can be further reduced to 10GB (INT8 ) and 6GB (INT4), allowing ChatGLM-6B to be deployed on consumer-grade graphics cards.
• Longer sequence length: Compared with GLM-10B (sequence length 1024), ChatGLM-6B has a sequence length of 2048, supporting longer conversations and applications.
• Human intention alignment training: Supervised Fine-Tuning, Feedback Bootstrap, Reinforcement Learning from Human Feedback and other methods are used to make The model is beginning to have the ability to understand the intent of human instructions. The output format is markdown for easy display.

However, due to the small capacity of the ChatGLM-6B model, there are inevitably some limitations and shortcomings, including:

• Relatively weak model memory and language capabilities. ChatGLM-6B may generate incorrect information when faced with many factual knowledge tasks, and it is not very good at solving logical problems (such as mathematics, programming).
• May generate harmful illustrations or biased content: ChatGLM-6B is only a preliminary language model aligned with human intent and may generate harmful or biased content.
• Weak multi-turn dialogue capability: ChatGLM-6B’s context understanding capability is not sufficient, and context loss may occur when faced with long answer generation and multi-turn dialogue scenarios. and misunderstanding situations.

#The GLM team stated that ChatGLM is still far behind the world's top large model research and products. In the future, it will continue to develop and open source updated versions of ChatGLM and related models. The GLM team also welcomes everyone to download ChatGLM-6B and conduct research and (non-commercial) application development based on it.

The above is the detailed content of ChatGLM, the Tsinghua-based Gigabit base conversation model, has launched internal testing and is an open source single-card version of the model.. For more information, please follow other related articles on the PHP Chinese website!