Hausa news topic classifier
Fine-tuned from Davlan/afro-xlmr-base on the Hausa portion of MasakhaNEWS (Adelani et al., 2023) to
predict one of 7 topics: business, entertainment, health, politics, religion, sports, technology.
Results (official MasakhaNEWS Hausa test split, 637 articles)
- Accuracy: 0.9058
- Macro-F1: 0.8981
Training
- Run: AfroXLMR-base | head256 | lr1e-05; max length 256 tokens; truncation: head
- Learning rate 1e-05, batch size 16 x 1; model selected on validation macro-F1.
Limitations
- Trained on BBC Hausa news only; expect lower accuracy on social media, other news sites, or other dialects/spelling conventions.
- Knows only the 7 topics above and has no "none of these" class.
- Small training set (about 2.2k articles); the test set has 637 articles, so small differences between models are within noise.
Intended use
Educational / research demonstration. Check the base model's card and the dataset card for license terms before any other use.
- Downloads last month
- 46