| { | |
| "rule": "flag when p(class = deliberately_malicious) >= class_malicious OR p(scope = yes) >= scope", | |
| "frozen": { | |
| "class_malicious": 0.9, | |
| "scope": 0.9, | |
| "fitted_on": "English/Python-heavy dev set, <= 5% false alarms; used for every result in the card" | |
| }, | |
| "multilingual_option": { | |
| "class_malicious": 0.9, | |
| "scope": 0.05, | |
| "fitted_on": "a 10% repo-level dev split of the 14-language set, <= 5% false alarms", | |
| "effect": "on the other 90%: recall 421/443 (95%) at 43/1900 (2.3%) real false alarms, vs 90% at 0.5% with the frozen thresholds" | |
| }, | |
| "advice": "fit thresholds on a sample of your own data (a few hundred benign items and a few dozen known-bad ones) for your false-alarm budget" | |
| } |