WebGPT: Improving the factual accuracy of language models through web browsing
OpenAI Blog
OpenAI fine-tuned GPT-3 to answer open-ended questions more accurately by giving it access to a text-based web browser. The model improved from 68% to 79% accuracy on a factual QA benchmark when equipped with browsing capability. This allows language models to retrieve current information and verify facts rather than relying solely on training data.
Why it matters
We’ve fine-tuned GPT-3 to more accurately answer open-ended questions using a text-based web browser.
