d1-3b-cli-command-risk / eval /guardrail.json
yannelli's picture
Fine-tuned checkpoint and model card
e1e980e verified
Raw History Blame Contribute Delete
902 Bytes
{
"cases": 12689,
"families": 97,
"level_macro_f1": 0.669,
"level_interval": [
0.484,
0.781
],
"level_accuracy": 0.756,
"f1_by_level": {
"safe": 0.989,
"low": 0.257,
"moderate": 0.86,
"high": 0.508,
"critical": 0.733
},
"category_macro_f1": 0.929,
"category_interval": [
0.724,
0.978
],
"f1_by_category": {
"data_destruction": 0.876,
"device_overwrite": 0.994,
"overbroad_permissions": 0.998,
"privilege_escalation": 1.0,
"remote_code_execution": 1.0,
"unsafe_download_exec": 1.0,
"resource_exhaustion": 1.0,
"credential_exposure": 1.0,
"irreversible_config": 0.693,
"history_log_tampering": 0.812,
"network_disruption": 1.0,
"signal_kill": 1.0,
"shell_footgun": 0.704
},
"risky_f1": 0.997,
"reversible_f1": 0.859,
"requires_root_f1": 0.911,
"cutoff_at_1pct_validation_fpr": 0.714,
"test_safe_fpr": 0.0168,
"recall_high_critical": 0.999
}