SWE-bench Pro seems to have a significant number of broken/disputable instances: https://openai.com/index/separating-signal-from-noise-coding-evaluations