Building better AI benchmarks: How many raters are enough?

Google Research · 3 months ago

Researchers investigated the optimal trade-off between the number of items rated and the number of raters per item in AI benchmarking, using simulations across subjective datasets like toxicity detection and chatbot safety evaluation. The standard practice of using 1 to 5 raters per item often proves insufficient, with more than 10 raters per item typically needed to capture human disagreement and achieve reproducible results. The findings show that practitioners can achieve highly reproducible benchmarks with approximately 1,000 total annotations by optimizing the ratings-per-item ratio based on their chosen metric, rather than pursuing the traditional single-truth paradigm.