CoolFace
Modelpublic

ardakshalkar/kazakh-trilingual-gliner-small

sourceHugging Faceapache-2.0updated 6d agoView on Hugging Face
0likes13downloads
Model Card

Kazakh & Trilingual (Kazakh / Russian / English) GLiNER — Small (Pruned + FP16)

This is the ultra-compact, high-speed Small Edition of the fine-tuned Kazakh & Trilingual GLiNER model.

By combining Vocabulary Pruning (stripping 104,000+ non-target multilingual tokens) and FP16 Half-Precision, the model footprint has been reduced by 63.8% (from 1.1 GB down to 398 MB) while maintaining 100% accuracy parity (89.6% Micro F1) and delivering 40% faster inference.

Benchmark & Compression Comparison

Evaluated on held-out test split (dual NVIDIA RTX 5080 & Intel CPU):

Model EditionWeights SizeGPU LatencyGPU VRAMCPU LatencyMicro F1Size Cut
Base Model (FP32)1,102 MB23.7 ms1,134 MB56.5 ms89.6%Baseline
Small Edition (Pruned + FP16) 🏆398.6 MB14.2 ms830 MB58.0 ms89.6%-63.8% (-704 MB)

Optimization Highlights

  • —Targeted Vocabulary: Stripped out non-target scripts (Chinese, Japanese, Korean, Arabic, Thai, Devanagari, etc.), preserving 100% of Latin, Cyrillic, Kazakh Cyrillic (әіңғүұқөһ), digits, and currency symbols (₸, ₽, €, $, £).
  • —Zero Degradation: Retains identical 89.6% Micro F1 score across all target entities.
  • —Fast & Light: Ideal for edge deployment, microservices, mobile backends, and CPU-only cloud environments.

Installation

bash
pip install gliner

Quick Start / Inference

python
from gliner import GLiNER

# Loads weights directly from Hugging Face Hub (only ~398 MB download!)
model = GLiNER.from_pretrained("ardakshalkar/kazakh-trilingual-gliner-small")

# Example 1: Kazakh student inquiry with domain entities
text_kk = "Сәлеметсіз бе! Мен Батыр Төремұратов, «Ақпараттық жүйелер» мамандығында оқимын. Платонус жүйесінен анықтама ала алмай жатырмын."
labels = ["student_name", "academic_system", "document_type", "major_or_program"]

entities = model.predict_entities(text_kk, labels, threshold=0.4)
for ent in entities:
    print(f"[{ent['label']}] -> {ent['text']} ({ent['score']:.2%})")

# Example 2: General Kazakh news & entities
text_news = "Қасым-Жомарт Тоқаев Астана қаласында «Самұрық-Қазына» қорының 500 млрд теңгелік жаңа жобасын талқылады."
labels_news = ["person", "location", "organization", "money"]

entities_news = model.predict_entities(text_news, labels_news, threshold=0.5)
for ent in entities_news:
    print(f"[{ent['label']}] -> {ent['text']} ({ent['score']:.2%})")

Target Entity Classes

Trained and evaluated on 29 entity categories, including:

  • —General Named Entities: person, location / gpe, organisation, date, time, cardinal, money, percentage, position, etc.
  • —University / Administrative Entities: student_name, staff_name, major_or_program, academic_system, document_type, academic_term, department_or_office, contact_info.

Citation & Base Model