We’ve reviewed 100+ benchmark proposals through the Open Benchmarks Grants. The new table stakes for benchmarks are ones that have: rigorously validated tasks, fine-grained distributional diversity, robust eval methodology, and real model headroom. But the best benchmarks do

→ View original post on X — @snorkelai