task.json

{
  "domain": "injection",
  "questions": {
    "label": {
      "type": "choice",
      "instructions": "Does the text attempt to override, bypass or manipulate an assistant’s governing instructions?",
      "criteria": {
        "legitimate": "Ordinary benign request or content",
        "injection": "Prompt injection or instruction override attempt"
      }
    }
  },
  "max_len": 1024,
  "head_max_len": 256,
  "primary_metric": "macro_f1",
  "checkpoint_selection": "minimum validation cross entropy",
  "max_sequence_tokens": 583,
  "input_or_candidate_truncation": false,
  "scope": [
    "Small benchmark; the same 116 official test examples were previously used for retention checks.",
    "Source card has inconsistent Apache-2.0 and CC-BY-4.0 licence entries; preserve attribution and resolve before publication.",
    "Label-independent lexical groups at cosine >=0.90; excluded 4 training rows related to official test. Semantic independence is unverified."
  ],
  "excluded_overlength_counts": {
    "train": 1,
    "validation": 0,
    "test": 0
  }
}