This model classifies text as either hate speech or not hate speech (binary classification). It was fine-tuned on the HateXplain dataset and exported to ONNX format with INT8 quantization for efficient on-device inference in browser extensions.

Training Data

Dataset: HateXplain (~20K samples)
Sources: Twitter, Gab
Label Mapping:
- hate (1): Hate speech
- not_hate (0): Offensive + Normal content merged

Evaluation Results

Metric	Score
F1 (Macro)	0.8302
Accuracy	0.8567
Precision	0.8342
Recall	0.8266

Model Comparison (RQ3)

Model	F1	Size (Quantized)
MiniLM (this)	0.830	32.6 MB
DistilBERT	0.825	64 MB
TinyBERT	0.816	15 MB

Intended Use

Primary: Browser extension for real-time hate speech detection
Framework: Transformers.js for on-device inference
Privacy: 100% client-side processing, no data sent to servers

Usage

With Transformers.js (Browser)

javascript

import { pipeline } from '@huggingface/transformers';

const classifier = await pipeline(
  'text-classification',
  'TaiwoOgun/minilm-hate-speech-onnx'
);

const result = await classifier('Some text to classify');
console.log(result);
// [{ label: 'hate', score: 0.85 }]

With Python

python

from transformers import pipeline

classifier = pipeline('text-classification', 'TaiwoOgun/minilm-hate-speech')
result = classifier('Some text to classify')

Limitations

Language: English only
Binary: Does not distinguish between types of hate speech
Context: May miss context-dependent hate speech or sarcasm
Dataset Bias: Trained on Twitter/Gab data, may not generalize to all platforms

Citation

If you use this model, please cite:

bibtex

@misc{ogunbanwo2025hatespeech,
  author = {Ogunbanwo, Taiwo},
  title = {Real-Time Hate Speech Detection Browser Extension},
  year = {2025},
  publisher = {HuggingFace},
  url = {https://huggingface.co/TaiwoOgun/minilm-hate-speech-onnx}
}

License

MIT License