2 ms·
Why do none of the benchmarks test for hallucinations?
by dom96 7mo ago
Why do none of the benchmarks test for hallucinations?
- netule 7mo ago[flagged]
- tedsanders 7mo agoIn the text, we did share one hallucination benchmark: Claim-level errors fell by 33% and responses with an error fell by 18%, on a set of error-prone ChatGPT prompts we collected (though of course the rate will vary a lot across different types of prompts). Hallucinations are the #1 problem with language models and we are working hard to keep bringing the rate down. (I work at OpenAI.)