banglabert-sentence-router-onnx

BanglaBERT Sentence Router (ONNX) - Shohoj Lipi 🧠

This is an ONNX-optimized, binary sentence classification model designed to detect complex Bengali sentences that are difficult for dyslexic readers (ages 6–14). It was created for Shohoj Lipi (সহজ লিপি), a project for the SciBlitz AI Challenge 2026 (Track C).

Model Details

• Base Model: csebuetnlp/banglabert (Electra architecture)
• Format: ONNX (CPU optimized for ultra-fast, low-memory inference)
• Task: Binary Classification (0 = Simple, 1 = Complex)
• Training Data: 96,000+ sentences from the BengaliReadability dataset

Intended Use

This model acts as a "Router" in a text simplification pipeline. It rapidly scans Bengali sentences and flags the complex ones (Label 1) to be sent to a Large Language Model (like Gemini) for simplification and syllable-scaffolding, while allowing simple sentences (Label 0) to pass through untouched.

Why ONNX?

Exporting BanglaBERT to ONNX reduces CPU inference time from ~2.0 seconds to ~300ms and cuts RAM usage by over 60%, allowing it to run smoothly on free-tier hosting services (like Render or Railway) without needing a GPU.

Usage (Python)

You can load this model locally using onnxruntime and HuggingFace's tokenizers:

import numpy as np
from transformers import AutoTokenizer
from optimum.onnxruntime import ORTModelForSequenceClassification

# Load tokenizer and ONNX model
tokenizer = AutoTokenizer.from_pretrained("YOUR_USERNAME/banglabert-sentence-router-onnx")
model = ORTModelForSequenceClassification.from_pretrained("YOUR_USERNAME/banglabert-sentence-router-onnx")

text = "বাংলাদেশের অর্থনীতি মূলত কৃষিনির্ভর।"
inputs = tokenizer(text, return_tensors="pt")

# Inference
outputs = model(**inputs)
logits = outputs.logits.detach().numpy()
predicted_class = np.argmax(logits, axis=1)[0]
confidence = np.max(np.exp(logits) / np.sum(np.exp(logits), axis=1, keepdims=True))

print("Class:", "Complex" if predicted_class == 1 else "Simple")
print("Confidence:", confidence)

Performance

• Validation Accuracy: 85.0%
• F1 Score: > 0.80

Limitations

• The model struggles with highly technical medical or legal jargon, as it was trained primarily on NCTB school textbooks and literature.
• It evaluates complexity at the sentence level, not the document level. It should be used in conjunction with a document-level readability heuristic.