zikabyte/garda-setfit

GARDA - SetFit Grooming Risk Classifier

Model Description

This is a SetFit model fine-tuned from sentence-transformers/all-MiniLM-L6-v2 to flag indications of online predatory grooming in English-language chat messages. It is a binary classifier (label 1 = suspicious, 0 = normal) trained on short conversation windows (a message plus the few messages before it, tagged self/other) rather than isolated single lines.

This model is a triage aid, not a verdict. See "Bias, Risks and Limitations" below before using it for anything beyond a demo/research context.

Training Data

Fine-tuned on the PAN-2012 Sexual Predator Identification corpus, a restricted-access, non-redistributable dataset of real chat logs used under permission for academic research. No excerpts of the raw dataset are included in this model card or repository - only the resulting model weights are shared, consistent with the dataset's usage terms.

Uses

Direct Use for Inference

from setfit import SetFitModel

model = SetFitModel.from_pretrained("zikabyte/garda-setfit")
probas = model.predict_proba([
    "self: hey\nother: hi\nother: how old are you?",
])