Binary TFGBV classifier fine-tuned on afro-xlmr-base with a linear classification head with sigmoid ≥ 0.5. Fine-tuning done on the combined benchmark + ALTO gold training data (meedan/alto 20260813 split, config sw), using the unredacted doxxing text for the items that are redacted in the public dataset.

The weights are a plain safetensors state dict for the BinaryClassifier module defined in the repo notebook. See that notebook for load_best_model and predict.