Seziberagabriel/kinyarwanda-news-classifier

Kinyarwanda News Topic Classifier

AfriBERTa (castorini/afriberta_base) fine-tuned to classify Kinyarwanda news articles into 14 topics: politics, sport, economy, health, entertainment, history, technology, tourism, culture, fashion, religion, environment, education, relationship.

Training data

KINNEWS (Niyongabo et al., 2020), kinnews_cleaned config. Duplicates and train/test overlaps were removed: 8,167 train / 908 validation / 2,008 test articles. Input text is lowercased title + content, truncated to 256 tokens.

Training

lr 3e-5, 3 epochs, batch size 16, max length 256, best checkpoint chosen by validation macro-F1.

Results (test set, 2,008 articles)

Accuracy 0.778, macro-F1 0.676. A TF-IDF + Logistic Regression baseline reached accuracy 0.802 and macro-F1 0.722 on the same split.

Limitations

News domain only; long articles are truncated; weak on rare classes such as history and environment; some articles (e.g. COVID-19 news) have ambiguous topic labels.

Code

https://github.com/Seziberagabriel/kinyarwanda-news-classifier