Eval anti-cheat idea: serve models a stale HF cache from before grader fixes

willcb · x · 2026-08-28

In a short exchange on eval contamination, willcb suggests giving models a stale HuggingFace cache from before the eval author inspected the data and patched grader bugs; the reply argues it'd be better to just turn off web access instead of praying. A quick debate on preventing benchmark leakage and cheating.

Original post →

More from coding & agent

coding & agent channel →