v3.1: fix residual false positives on short bare commands (pwd, whoami, echo X, cat X) — benign-command acc 0.93->0.98, realistic metrics unchanged (macro-F1 0.931, malicious recall 0.945) fb1efb9 verified sam-at-axiotic commited on Jun 9
Model card: full v3.1 shell-command stats (per-cell table + SmolVM worked example) b28d8c0 verified sam-at-axiotic commited on Jun 9
v3.1: fix residual false positives on short bare commands (pwd, whoami, echo X, cat X) — benign-command acc 0.93->0.98, realistic metrics unchanged (macro-F1 0.931, malicious recall 0.945) c476867 verified sam-at-axiotic commited on Jun 9
v3: fix shell-command false positives (persona command/code datagen + class weighting); realistic macro-F1 0.926->0.931, benign-command acc 0.05->0.93, malicious recall 0.923->0.945 6dc5c44 verified sam-at-axiotic commited on Jun 9
ogma-prompt-injection v2: generated-benign retrain, label smoothing, RoPE fix baked in 2475d58 verified sam-at-axiotic commited on Jun 3