Commit History

12 days ago

  1. v2.1: recommended threshold 0.3 -> 0.05 (re-checked on the released seed, no test data); seed and soup results

    Verified signature
    0e6a8e7
  2. Add recall mode (v1 + v2 average) to detect.py; link the false-alarm benchmark

    Verified signature
    de8db85
  3. Release v2: mmBERT-base, obfuscation augmentation, decoder; selection protocol kept apart from test

    Verified signature
    6f98f61
  4. Release v1: Turkish prompt injection classifier (BERTurk, source-held-out evaluation)

    Verified signature
    d60fcbc