Mirror of BultenHub/bhkai, BültenHub's official repository.

BHKAI: Turkish news category classifier

BHKAI assigns a category to each incoming story. When it is confident it answers on its own; otherwise the story goes to the AI model.

Model

Method: TF-IDF (300,000 features, min_df 3) with an SGD classifier (alpha 2e-6).

Version: bhkai-20261030, retrained nightly and promoted automatically after its exam.

Training data: 437,769 stories labelled by the teacher model (kategori_kaynagi='ai'). Data cut-off: 2026-10-30 03:24.

Exam

The exam uses the last 2 days: 749 stories, half of them, split by crc32.

Threshold	Coverage	Agreement
0.8 (live)	0.512	0.997
0.7	0.659	0.973
0.5	0.855	0.902

The previous live version scored coverage 0.488 and agreement 1.000 at threshold 0.8.

Files

• agirlik.npz: weights
• sozluk.json: vocabulary
• bilgi.json: version, exam and comparison details

They are loaded by BültenHub's pipeline/bhkai_* code.

License

CC BY-NC 4.0 (non-commercial). For commercial use, contact https://bultenhub.com.