ComputeBench: Instruction-following benchmarks for long, step-by-step arithmetic(notdian.github.io)

1 ポイント投稿者 notdian7 か月前1 コメント

notdian.github.io

ComputeBench: Instruction-following benchmarks for long, step-by-step arithmetic

https://notdian.github.io/computebench/

1 コメント

notdian7 か月前

Vibecoded this after seeing models do amazing things but still drift on simple recursive steps; tracks exact match, answer accuracy, prefix correctness. Feedback welcome.