WebGLM

Why we included this project

WebGLM combines a large language model with live web search, so it can pull up current pages and answer questions with sources instead of falling back on training data alone. The system retrieves relevant content, drafts an answer, then uses a scorer trained on human preferences to filter out responses that aren't actually useful. You can run it as a command-line tool or as a small web service, with SerpAPI or Bing for search and 2B or 10B model sizes to fit your hardware. For teams prototyping grounded question answering that needs up-to-date answers, it's a clear example of how retrieval, generation, and preference scoring can work together. It also includes training code and the WebGLM-QA dataset if you want to reproduce or adapt the method.