haithamd/fake-news-bert

Fake News BERT (8-bit ONNX)

bert-base-uncased fine-tuned to classify news text as Fake or Real, from my MSc dissertation Fake News Detection Using Transformer-Based and Traditional NLP Models (MSc Data Science and Artificial Intelligence, University of Liverpool). Exported to ONNX with 8-bit dynamic quantisation so it runs in the browser with Transformers.js.

Try it: Fake News Detector

Training

• Data: Kaggle Fake and Real News dataset, 44,673 articles after de-duplication, split 80/10/10 (stratified, seed 42).
• Input: title + body, cleaned (HTML/URLs removed, lower-cased, stop words and non-alphabetic tokens dropped, WordNet-lemmatised), truncated to 256 tokens.
• 4 epochs, AdamW, learning rate 2e-5, batch size 16, on a Colab T4 GPU.

Results

On the dissertation's held-out test set (4,468 articles): accuracy 0.9935, F1 0.9931 for this 8-bit model, compared with F1 0.9936 for the full-precision model.

Usage

The model expects text cleaned the same way as in training; see model.js in the Space for the exact steps.

import { pipeline } from "@huggingface/transformers";
const classify = await pipeline("text-classification", "haithamd/fake-news-bert", { dtype: "q8" });
await classify("cleaned text here", { top_k: null }); // [{ label: "Real", score }, { label: "Fake", score }]

Limitations

The model judges writing style, not truth. About 99% of the "Real" training articles are Reuters wire copy, and all come from 2016–2017, mostly US politics. A false claim written in a neutral wire-service style can be classified as Real, and short texts such as tweets are less reliable than full articles.