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

BHKAI: Turkish news category classifier

BHKAI («BültenHub Kategori AI») assigns a category to each incoming Turkish news story. In BültenHub it answers on its own when its confidence is at least 0.8; otherwise the story goes to a larger teacher model. It is retrained every night and replaces the production version only if it passes the same exam.

Version

Version	bhkai-20261112
Trained	2026-11-12
Method	TF-IDF word 1–2-grams (300,000 features, min_df 3) + SGD logistic classifier (alpha 2e-06)
Training data	553,578 stories labelled by the teacher model (data cut-off 2026-11-12)
Replaced	bhkai-20261110

Exam (the metrics the nightly gate uses)

Held-out recent stories the previous version never saw (10968 stories; stories the production model was confident on are weighted by 1/audit-rate so the exam matches production traffic). Coverage = share of stories BHKAI decides on its own; agreement = same category as the teacher model.

Confidence ≥	Coverage	Agreement	Previous: coverage	Previous: agreement
0.5	0.838	0.881	0.824	0.872
0.7	0.613	0.948	0.595	0.947
0.8 (production)	0.481	0.969	0.478	0.965
0.85	0.345	0.988	0.299	0.990
0.9	0.212	1.000	0.202	1.000

Gate decision: promoted — at confidence ≥ 0.8: coverage 48.1% (was 47.8%), agreement with teacher 96.9% (was 96.5%).

Usage (Türkçe örnek)

# pip install huggingface_hub numpy
import json, sys
from huggingface_hub import snapshot_download

yol = snapshot_download("BultenHub/bhkai")
sys.path.insert(0, yol)
from bhkai import Bhkai

model = Bhkai(yol)
kategori_id, guven = model.tahmin(
    "Kurgukent'te 24 saat açık kent kütüphanesi hizmete girdi",           # başlık (kurgusal)
    "Kurgukent Belediyesi'nin eski itfaiye binasında kurduğu kütüphane dün açıldı. "
    "18 bin kitap ve 120 kişilik çalışma salonu bulunuyor.",              # metnin başı
    "Kurgukent Haber",                                                    # kaynak adı
)
adlar = json.load(open(f"{yol}/kategoriler.json", encoding="utf-8"))
print(adlar.get(str(kategori_id), kategori_id), round(guven, 2))
# BültenHub'da güven ≥ 0.8 ise kararı BHKAI verir, değilse haber büyük modele gider.

Files: agirlik.npz (weights), sozluk.json (vocabulary), bilgi.json (version, exam), bhkai.py (numpy inference), kategoriler.json (category id → name).

Limitations

• Labels come from a teacher LLM, not from human editors; agreement is measured against that teacher.
• Trained on Turkish news headlines and the first 500 characters of the text; other domains and languages are out of scope.
• Low-confidence predictions are unreliable by design: use the confidence threshold.

Changelog

• 2026-11-12 bhkai-20261112: at confidence ≥ 0.8: coverage 48.1% (was 47.8%), agreement with teacher 96.9% (was 96.5%)
• 2026-11-10 bhkai-20261110: at confidence ≥ 0.8: coverage 45.2% (was 44.4%), agreement with teacher 97.4% (was 97.3%)
• 2026-11-04 bhkai-20261104: at confidence ≥ 0.8: coverage 57.2% (was 57.7%), agreement with teacher 97.4% (was 97.2%)
• 2026-11-03 bhkai-20261103: at confidence ≥ 0.8: coverage 46.3% (was 46.3%), agreement with teacher 97.2% (was 95.3%)
• 2026-11-01 bhkai-20261101: at confidence ≥ 0.8: coverage 56.1% (was 56.1%), agreement with teacher 96.5% (was 96.7%)

License

CC BY-NC 4.0 (non-commercial). The base model, if any, keeps its own license. For commercial use, contact https://bultenhub.com.