BERT fine-tuned on AG News (Topic Classification)
Model description
This model is a fine-tuned version of bert-base-uncased for topic classification on the AG News dataset. It classifies English news articles into four categories: World, Sports, Business, and Sci/Tech.
Training data
- Dataset: fancyzhx/ag_news
- Training examples: 108000 (after reserving 12000 for validation)
- Validation examples: 12000
- Test examples (official, held out for final evaluation only): 7600
- Classes: World, Sports, Business, Sci/Tech
Training procedure
- Base model:
bert-base-uncased - Adaptation method: Full fine-tuning
- Encoder learning rate: 2e-5
- Classification head learning rate: 1e-3
- Train batch size: 16, eval batch size: 32
- Epochs: 1
- Seed: 42
Evaluation results (official AG News test set)
- Accuracy: 0.9445
- Macro F1: 0.9444
- Official test loss: 0.1754
Intended use
Automatic topic classification of English news articles into the four AG News categories. Can serve as a baseline for organizing or filtering English news content by topic.
Limitations
- Trained and evaluated only on AG News; may not generalize to other domains, writing styles, or languages.
- English only.
- Input truncated to 128 tokens.
- Fine-tuned for a single epoch.
References
- Devlin, J., Chang, M.-W., Lee, K., & Toutanova, K. (2019). BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding.
- Zhang, X., Zhao, J., & LeCun, Y. (2015). Character-level Convolutional Networks for Text Classification (source of the AG News corpus).
- Hugging Face Transformers documentation: https://huggingface.co/docs/transformers