Add safeagents judge bench results
#4
by adaamko - opened
our safeagents judge bench v0 run, test split, unmodified benchmark runner, fixed 0.5 cut-off, no tuning. revision bbc0206efbbb
balanced accuracy 0.716, AUROC 0.952, false-block rate 0.564, category macro F1 0.359
the gated adapters stay off for these questions, so the pinned base d3-mini@566da0ea answers
Xunzhuo changed pull request status to merged