I have spent some time doing this for these benchmarks — the model still does make mistakes. Of the questions I can understand, (roughly half in this case) abou

I have spent some time doing this for these benchmarks — the model still does make mistakes. Of the questions I can understand, (roughly half in this case) about half were real errors and half were broken questions.