CoolFace
Modelpublic

kanekoyuichi/promptgate-classifier-v2

sourceHugging Facemitupdated 5mo agoView on Hugging Face
0likes12downloads
Model Card

promptgate-classifier-v2

Fine-tuned DistilBERT sequence classifier for PromptGate prompt-injection screening.

Labels:

  • —SAFE: benign input
  • —ATTACK: prompt-injection-like input

This model is intended for use through PromptGate:

python
from promptgate import PromptGate

gate = PromptGate(detectors=["rule", "classifier"])
result = gate.scan("Ignore all previous instructions.")

Reference holdout results used during development:

DetectorRecallSpecificityPrecisionAccuracy
classifier v2 @ 0.592.5%85.0%86.0%88.8%

These numbers are reference values for a fixed development holdout and are not a guarantee of production performance.