LiveBench is a benchmark for LLMs designed with test set contamination and objective evaluation in mind. It has the following properties:
LiveBench is designed to limit potential contamination by releasing new questions monthly, as well as having questions based on recently-released datasets, arXiv papers, news articles, and IMDb movie synopses.
Each question has verifiable, objective ground-truth answers, allowing hard questions… See the full description on the dataset page:
https://huggingface.co/datasets/livebench/instruction_following.
