TinyBanglaClickbaitBERT

A compact Bengali clickbait detection model distilled from BanglaClickbaitBERT.

Model Details

• Architecture: 4-layer Transformer (384-d, 6 heads)
• Parameters: 20.1M (vs 110.2M teacher → 5.5x compression)
• Tasks: Binary (clickbait/non-clickbait) + 11-class multiclass

Performance

Task	Student F1	Teacher F1	Retention
Binary (macro)	0.8543	0.8849	96.5%
Multiclass (macro)	0.4201	0.5088	82.6%

Speed

• Teacher: 7.9 ms | Student: 3.4 ms | 2.3x speedup
• Teacher: 440.9 MB | Student: 80.3 MB | 5.5x compression

Usage

import torch
checkpoint = torch.load("tiny_bangla_clickbait_bert.pt")
# See label_config.json for class mappings

Distilled from

khalidsyfullah/bangla-clickbait-multitask