lanretto/naijaweb-quality-classifier

NaijaWeb Quality Classifier

A binary text-quality classifier (keep/drop) for Hausa, Igbo, and Yoruba web text, fine-tuned from Davlan/afro-xlmr-base for filtering CommonCrawl pretraining data.

Training data

Trained on paid, agreement-audited human annotations from idealab-cs2/naijaweb-annotations-trained.

Evaluation (internal held-out split, host-balanced)

Metric	Score
Accuracy	0.713
F1	0.752
Precision	0.670
Recall	0.855

Note: provided test split numbers are known to be host-skewed (see dataset card) and reported separately as diagnostic only.