wolf-defender-jev / eval /heldout_benchmarks.json
domi29's picture
Card: MIT licence, no training section
1849983 verified
Raw History Blame Contribute Delete
10 kB
{
"model": "Wolf Defender Jev",
"description": "33 held-out benchmarks, primary decision each; the 10 Japanese indirect-injection probe sets count as one. 'external' = source not seen during training.",
"macro_all_33": {
"balanced_acc": 0.9276,
"auroc": 0.9556,
"tpr@1%fpr": 0.7477,
"fpr@0.5": 0.0534,
"ece": 0.0596
},
"macro_external_10": {
"balanced_acc": 0.8481,
"auroc": 0.8645,
"tpr@1%fpr": 0.3084,
"fpr@0.5": 0.1298,
"ece": 0.1315
},
"capabilities": {
"Prompt injection (direct + indirect)": {
"benchmarks": 14,
"balanced_acc": 0.952,
"auroc": 0.9905,
"tpr@1%fpr": 0.9208,
"fpr@0.5": 0.0179,
"ece": 0.0412
},
"Agent / tool security": {
"benchmarks": 6,
"balanced_acc": 0.9632,
"auroc": 0.9885,
"tpr@1%fpr": 0.9293,
"fpr@0.5": 0.0122,
"ece": 0.028
},
"Sensitive data / PII": {
"benchmarks": 4,
"balanced_acc": 0.9802,
"auroc": 0.9947,
"tpr@1%fpr": 0.9675,
"fpr@0.5": 0.0046,
"ece": 0.0076
},
"Multilingual / non-English": {
"benchmarks": 8,
"balanced_acc": 0.8975,
"auroc": 0.959,
"tpr@1%fpr": 0.7535,
"fpr@0.5": 0.033,
"ece": 0.0813
},
"Jailbreak / unsafe requests": {
"benchmarks": 6,
"balanced_acc": 0.8335,
"auroc": 0.8881,
"tpr@1%fpr": 0.3604,
"fpr@0.5": 0.1712,
"ece": 0.1407
},
"Hard negatives / over-defense (benign-only; FPR)": {
"benchmarks": 6,
"balanced_acc": 0.944,
"auroc": 0.8346,
"tpr@1%fpr": 0.18,
"fpr@0.5": 0.0711,
"ece": 0.0487
}
},
"benchmarks": {
"aegis2-content-safety-test": {
"name": "Aegis 2.0 (test)",
"decision": "harmful.request",
"n": 1750,
"external": false,
"balanced_acc": 0.8445,
"auroc": 0.9168,
"tpr@1%fpr": 0.3709,
"fpr@0.5": 0.1683,
"ece": 0.1121
},
"agentic-pi-5k-test": {
"name": "agentic-PI-5k (test)",
"decision": "injection.presence",
"n": 1365,
"external": false,
"balanced_acc": 0.9829,
"auroc": 0.9995,
"tpr@1%fpr": 0.9857,
"fpr@0.5": 0.0377,
"ece": 0.0176
},
"arabic-guardrail-hard": {
"name": "Arabic guardrail (hard)",
"decision": "harmful.request",
"n": 1456,
"external": false,
"balanced_acc": 0.9808,
"auroc": null,
"tpr@1%fpr": null,
"fpr@0.5": 0.0192,
"ece": 0.0138
},
"boundary-pairs-test": {
"name": "boundary pairs (test)",
"decision": "injection.presence",
"n": 879,
"external": false,
"balanced_acc": 1.0,
"auroc": 1.0,
"tpr@1%fpr": 1.0,
"fpr@0.5": 0.0,
"ece": 0.0
},
"browsesafe-test": {
"name": "BrowseSafe (test)",
"decision": "injection.presence",
"n": 8227,
"external": false,
"balanced_acc": 0.8273,
"auroc": 0.9343,
"tpr@1%fpr": 0.6393,
"fpr@0.5": 0.0329,
"ece": 0.129
},
"deepset-prompt-injections-multilingual-test": {
"name": "deepset prompt-injections multilingual (test)",
"decision": "injection.presence",
"n": 1322,
"external": false,
"balanced_acc": 0.9846,
"auroc": 0.9999,
"tpr@1%fpr": 1.0,
"fpr@0.5": 0.0,
"ece": 0.0076
},
"deepset-prompt-injections-test": {
"name": "deepset/prompt-injections (test)",
"decision": "injection.presence",
"n": 182,
"external": false,
"balanced_acc": 1.0,
"auroc": 1.0,
"tpr@1%fpr": 1.0,
"fpr@0.5": 0.0,
"ece": 0.0026
},
"gretel-pii-finance-multilingual-test": {
"name": "Gretel PII finance multilingual (test)",
"decision": "secret.presence",
"n": 3566,
"external": false,
"balanced_acc": 0.9341,
"auroc": 0.9805,
"tpr@1%fpr": 0.8812,
"fpr@0.5": 0.0129,
"ece": 0.0192
},
"gretel-pii-masking-en-test": {
"name": "Gretel PII masking (test)",
"decision": "secret.presence",
"n": 3512,
"external": false,
"balanced_acc": 0.9979,
"auroc": 1.0,
"tpr@1%fpr": 1.0,
"fpr@0.5": 0.0041,
"ece": 0.0045
},
"jackhhao-jailbreak-classification-test": {
"name": "jackhhao jailbreak-classification (test)",
"decision": "jailbreak.presence",
"n": 800,
"external": true,
"balanced_acc": 0.9129,
"auroc": 0.9511,
"tpr@1%fpr": 0.5887,
"fpr@0.5": 0.0772,
"ece": 0.0859
},
"jbb-behaviors": {
"name": "JailbreakBench behaviors",
"decision": "harmful.request",
"n": 200,
"external": true,
"balanced_acc": 0.745,
"auroc": 0.8404,
"tpr@1%fpr": 0.29,
"fpr@0.5": 0.34,
"ece": 0.2075
},
"jpi-ja": {
"name": "Japanese indirect injection (10 probe sets, mean)",
"decision": "injection.presence",
"n": 180,
"external": true,
"balanced_acc": 0.6817,
"auroc": null,
"tpr@1%fpr": null,
"fpr@0.5": null,
"ece": 0.2903
},
"lakera-gandalf-ignore-instructions-test": {
"name": "Lakera Gandalf ignore-instructions (test)",
"decision": "injection.presence",
"n": 328,
"external": false,
"balanced_acc": 1.0,
"auroc": null,
"tpr@1%fpr": null,
"fpr@0.5": null,
"ece": 0.0
},
"lakera-gandalf-summarization-test": {
"name": "Lakera Gandalf summarization (test)",
"decision": "injection.presence",
"n": 30,
"external": false,
"balanced_acc": 1.0,
"auroc": null,
"tpr@1%fpr": null,
"fpr@0.5": null,
"ece": 0.0
},
"lakera-mosscap-test": {
"name": "Lakera Mosscap (test)",
"decision": "prompt_leak.attempt",
"n": 6767,
"external": false,
"balanced_acc": 1.0,
"auroc": null,
"tpr@1%fpr": null,
"fpr@0.5": null,
"ece": 0.0
},
"massive-benign-test": {
"name": "MASSIVE benign (FP control)",
"decision": "injection.presence",
"n": 13042,
"external": false,
"balanced_acc": 1.0,
"auroc": null,
"tpr@1%fpr": null,
"fpr@0.5": 0.0,
"ece": 0.0
},
"nemotron-pii-test": {
"name": "Nemotron PII (test)",
"decision": "secret.presence",
"n": 7242,
"external": false,
"balanced_acc": 0.9921,
"auroc": 0.9984,
"tpr@1%fpr": 0.9909,
"fpr@0.5": 0.0014,
"ece": 0.0039
},
"nemotron-safety-guard-v3-test": {
"name": "Nemotron Safety Guard v3 (test)",
"decision": "harmful.request",
"n": 6914,
"external": false,
"balanced_acc": 0.7505,
"auroc": 0.8442,
"tpr@1%fpr": 0.1919,
"fpr@0.5": 0.0993,
"ece": 0.2029
},
"notinject-notinject-one": {
"name": "NotInject-1 (over-defense)",
"decision": "injection.presence",
"n": 226,
"external": true,
"balanced_acc": 0.9823,
"auroc": null,
"tpr@1%fpr": null,
"fpr@0.5": 0.0177,
"ece": 0.0144
},
"notinject-notinject-three": {
"name": "NotInject-3 (over-defense)",
"decision": "injection.presence",
"n": 226,
"external": true,
"balanced_acc": 0.9646,
"auroc": null,
"tpr@1%fpr": null,
"fpr@0.5": 0.0354,
"ece": 0.039
},
"notinject-notinject-two": {
"name": "NotInject-2 (over-defense)",
"decision": "injection.presence",
"n": 226,
"external": true,
"balanced_acc": 0.9912,
"auroc": null,
"tpr@1%fpr": null,
"fpr@0.5": 0.0088,
"ece": 0.0076
},
"or-bench-hard": {
"name": "OR-Bench hard (over-refusal)",
"decision": "harmful.request",
"n": 1319,
"external": false,
"balanced_acc": 0.9871,
"auroc": null,
"tpr@1%fpr": null,
"fpr@0.5": 0.0129,
"ece": 0.0074
},
"patronus-holdout-llmail-holdout": {
"name": "Patronus holdout: llmail-holdout",
"decision": "injection.presence",
"n": 5721,
"external": false,
"balanced_acc": 0.9997,
"auroc": 1.0,
"tpr@1%fpr": 1.0,
"fpr@0.5": 0.0,
"ece": 0.0008
},
"patronus-holdout-mosscap-unseen-lang": {
"name": "Patronus holdout: mosscap-unseen-lang",
"decision": "prompt_leak.attempt",
"n": 22,
"external": false,
"balanced_acc": 1.0,
"auroc": null,
"tpr@1%fpr": null,
"fpr@0.5": null,
"ece": 0.0
},
"patronus-holdout-synth-iid": {
"name": "Patronus holdout: synth-iid",
"decision": "injection.presence",
"n": 19494,
"external": false,
"balanced_acc": 0.969,
"auroc": 0.9975,
"tpr@1%fpr": 0.9509,
"fpr@0.5": 0.0026,
"ece": 0.0202
},
"patronus-holdout-synth-toolpoison-probe": {
"name": "Patronus holdout: synth-toolpoison-probe",
"decision": "tool_metadata.poisoning",
"n": 2509,
"external": false,
"balanced_acc": 1.0,
"auroc": 1.0,
"tpr@1%fpr": 1.0,
"fpr@0.5": 0.0,
"ece": 0.0
},
"polyguard-prompts-test": {
"name": "PolyGuard prompts (multilingual, test)",
"decision": "harmful.request",
"n": 64170,
"external": true,
"balanced_acc": 0.7157,
"auroc": 0.8318,
"tpr@1%fpr": 0.1749,
"fpr@0.5": 0.0778,
"ece": 0.211
},
"privy-pii-test": {
"name": "Privy PII (test)",
"decision": "sensitive.presence",
"n": 2383,
"external": false,
"balanced_acc": 0.9968,
"auroc": 0.9999,
"tpr@1%fpr": 0.9978,
"fpr@0.5": 0.0,
"ece": 0.0029
},
"simsonsun-jailbreak-prompts-dataset-1": {
"name": "simsonsun jailbreak prompts 1",
"decision": "jailbreak.presence",
"n": 67,
"external": true,
"balanced_acc": 0.8612,
"auroc": null,
"tpr@1%fpr": null,
"fpr@0.5": null,
"ece": 0.1411
},
"simsonsun-jailbreak-prompts-dataset-2": {
"name": "simsonsun jailbreak prompts 2",
"decision": "jailbreak.presence",
"n": 2284,
"external": true,
"balanced_acc": 0.8871,
"auroc": null,
"tpr@1%fpr": null,
"fpr@0.5": null,
"ece": 0.0945
},
"turkish-conversation-pi-test": {
"name": "Turkish conversation PI (test)",
"decision": "injection.presence",
"n": 243,
"external": false,
"balanced_acc": 0.9167,
"auroc": 0.9907,
"tpr@1%fpr": 0.8,
"fpr@0.5": 0.0,
"ece": 0.0632
},
"turkish-pi-1k-test": {
"name": "Turkish PI 1k (test)",
"decision": "injection.presence",
"n": 246,
"external": false,
"balanced_acc": 0.9664,
"auroc": 0.9922,
"tpr@1%fpr": 0.9114,
"fpr@0.5": 0.0882,
"ece": 0.0452
},
"xstest": {
"name": "XSTest (over-refusal)",
"decision": "harmful.request",
"n": 450,
"external": true,
"balanced_acc": 0.739,
"auroc": 0.8346,
"tpr@1%fpr": 0.18,
"fpr@0.5": 0.352,
"ece": 0.224
}
}
}