liar-roberta
RoBERTa-base fine-tuned on the LIAR dataset for binary fake-news detection — 64.5% accuracy, 0.635 macro-F1, beating a TF-IDF + LogReg baseline (0.603). Includes class-weighted training, calibration analysis (ECE 0.042), per-class error breakdown, threshold selection, ONNX export, and a live Spaces demo. Reproducible from one Colab notebook.
- 1
- MIT
- 4 wk ago
