BERT News Classifier
Predicts Business, Sports, or Technology from English news. Technology is AG News Sci/Tech, including science. World is excluded. Trained on 9000 balanced examples for 1 epoch(s). Validation rows: 900. Held-out test rows: 5699. Test accuracy: 0.9340. Test macro F1: 0.9340.
Architecture: bert-base-uncased final CLS embedding, dropout, Linear(768, 3). Preprocessing: lowercase, HTML/URL/noise cleanup, maximum 128 tokens. Cross entropy on logits during training; softmax at inference. This repository uses the custom NewsClassifier class in news_model.py; load with NewsClassifier.from_pretrained, not the standard AutoModel classifier. The companion app.py shows the complete loading and preprocessing workflow.
Limitations: English only; old news training data; three forced categories; long text is truncated; softmax scores are not calibrated confidence. See metrics.json for results and training settings.
- Downloads last month
- 14