Introduction to AutoWebGLM
AutoWebGLM is an exciting project aimed at revolutionizing web navigation through advanced language model technology. It's built on the substantial foundations of the ChatGLM3-6B model, enhancing its ability to streamline web browsing and improve the overall efficiency of navigating the internet. The project's main goal is to tackle everyday browsing challenges more effectively, offering a seamless web experience for users.
Key Features
-
HTML Simplification Algorithm: This innovative feature takes inspiration from how humans browse the web. It simplifies HTML content, making web pages easier for large language model (LLM) agents to understand without losing important information.
-
Hybrid Human-AI Training: The project employs a unique blend of human expertise and AI capabilities to build a rich set of web browsing data. This approach is crucial for curriculum training, significantly boosting the model's practical navigation skills.
-
Reinforcement Learning and Rejection Sampling: To enhance the comprehension and operational capabilities of the model, the project uses techniques like reinforcement learning and rejection sampling. These methodologies help break down complex tasks and improve interaction efficiency on web pages.
-
Bilingual Web Navigation Benchmark: AutoWebGLM introduces AutoWebBench, a robust tool designed for real-world web browsing tasks. This bilingual benchmark, supporting both Chinese and English, serves as a testing ground for refining and assessing the efficiency of AI web navigation agents.
Evaluation
The project has publicly shared its evaluation code, data, and testing environment for interested parties to explore and experiment with.
-
AutoWebBench & Mind2Web: These are the project’s evaluation datasets. They are available for download and use to measure the model’s performance. The inference code is accessible through the ChatGLM3-6B link, and users can calculate scores with a simple command-line execution.
-
WebArena and MiniWob++: AutoWebGLM has tailored these environments to fit its system's interaction needs, providing detailed modification and execution instructions in their respective resources.
Licensing and Contributions
AutoWebGLM is shared under the Apache-2.0 License, and all available data are intended solely for research purposes. The development team encourages users who find the project beneficial to support further advancement by starring the repository.
In academic circles, if the project contributes to research endeavors, citations following the provided format are appreciated.
In essence, AutoWebGLM symbolizes a forward-thinking approach to improving web interactions using AI, aiming to simplify and enhance user experience in our increasingly digital world.