SimpleBench
What it measures and how
Distinguish multiple-choice and open responses, fixing AVG@5, temperature and prompt protocol. A model’s leaderboard entry date is not its task evaluation date.
How this evidence is used
Candidate under observation: new-model coverage is timely. Official score reuse terms and a description of reproducible data versions are pending. Code licensing is not substituted; no automated collection or ranking yet.
Limitations and attribution
Everyday common-sense tasks offer complementary evidence, but task representativeness and multiple-choice versus open-response conditions affect interpretation.
Data license: Sample tasks and evaluation code are MIT. Display terms for the latest score script published separately on the official site remain unconfirmed.
Results are published by SimpleBench. Raw scores and the e/acc consensus index use different scales and cannot be added directly.