Are you looking for a lightweight and open-source alternative to ChatGPT for running and fine-tuning LLMs on your machine? Look no further than ChatGLM-6B! In recent weeks, several open-source ChatGPT alternatives have become popular, and in this article, we’ll explore the ChatGLM series and specifically ChatGLM-6B.
What is ChatGLM?
ChatGLM is a bilingual large-scale language model trained in both Chinese and English. Developed by researchers at Tsinghua University in China, the ChatGLM series is comparable in performance to other models such as GPT-3 and BLOOM. The following models are currently available :
- ChatGLM-130B: Open source LLM.
- ChatGLM-100B: Not open-sourced, but available by invitation-only access.
- ChatGLM-6B: a lightweight open source alternative.
These models appear similar to the GPT Group’s large-scale language models, but the General Language Model (GLM) pre-training framework makes them stand out.
How does ChatGLM work?
In machine learning, a GLM is a generalized linear model. In the case of ChatGLM, however, GLM stands for General Language Model. The objective optimization is formulated as an autoregressive space-filling problem, whereby spans of consecutive text are blanked and the text is sequentially reconstructed through these blanks. In addition, there is a long mask that randomly removes long blank text at the end of sentences to improve natural language understanding.
Another difference between ChatGLM and GPT-style models is in the type of attention used. While GPT models use unidirectional attention, ChatGLM models use bidirectional attention, which can capture dependencies better and improve performance on natural language understanding tasks.
ChatGLM-6B
ChatGLM-6B is a lightweight alternative model with approximately 6.2 billion parameters. The model is pre-trained with 1 trillion tokens, equally split between English and Chinese.
Pros and Cons of ChatGLM
ChatGLM-6B has the following advantages :
- Bilingual model optimized for user devices, so it can run with as little as 6 GB of memory.
- Good performance in both English and Chinese, making it a good choice for those working with both languages.
- Ability to perform well on a variety of tasks, including summarization and single– and multi-query chats.
ChatGLM-6B also has the following limitations :
- Poor performance in English, possibly due to the fact that most of the instructions used in training are in Chinese.
- ChatGLM-6B has significantly fewer parameters than other LLMs such as BLOOM, GPT-3, and ChatGLM-130B, which may result in poor performance when contexts are too long.
- Limited memory capacity may cause performance degradation in multi-turn chats.
Conclusion
ChatGLM-6B is a great lightweight and open-source alternative to ChatGPT for those looking to run and fine-tune LLMs locally. You can run ChatGLM-6B locally or try out the demo on HuggingFace spaces ( https://huggingface.co/spaces/multimodalart/ChatGLM-6B ).
