models/metadata.json

{
  "model": "microsoft/deberta-v3-small",
  "max_length": 256,
  "version": "3.0.0",
  "num_labels": 6,
  "labels": [
    "safe",
    "role_override",
    "goal_hijacking",
    "context_poisoning",
    "tool_manipulation",
    "cascading_amplification"
  ],
  "label2id": {
    "safe": 0,
    "role_override": 1,
    "goal_hijacking": 2,
    "context_poisoning": 3,
    "tool_manipulation": 4,
    "cascading_amplification": 5
  },
  "id2label": {
    "0": "safe",
    "1": "role_override",
    "2": "goal_hijacking",
    "3": "context_poisoning",
    "4": "tool_manipulation",
    "5": "cascading_amplification"
  },
  "augmentations": [
    "base64",
    "cyrillic_homoglyphs",
    "rot13",
    "zwsp"
  ],
  "schema": "6-class",
  "classes": {
    "0": "safe \u2014 no injection detected",
    "1": "role_override \u2014 DAN, persona hijacking, direct jailbreaks",
    "2": "goal_hijacking \u2014 hidden malicious secondary objective",
    "3": "context_poisoning \u2014 false authority/privilege in retrieved context",
    "4": "tool_manipulation \u2014 malicious tool call parameters",
    "5": "cascading_amplification \u2014 multi-stage escalation attacks"
  }
}