DistilBERT Fake News Classifier

A fine-tuned DistilBERT model that classifies news articles as Real or Fake.

Model Details

Model Description

This model is a fine-tuned version of distilbert-base-uncased for binary classification of news articles as Real or Fake. It was fine-tuned on the WELFake dataset after deduplication (removing ~9,337 duplicate rows) to prevent train/test leakage and ensure an honest evaluation.

• Developed by: Amey Dande
• Model type: Text classification (binary)
• Language(s) (NLP): English
• License: MIT
• Finetuned from model: distilbert-base-uncased

Model Sources

• Repository: https://github.com/AmeyDande/Fake_News_Classisier

Uses

Direct Use

This model can classify English-language news article text or headlines as real or fake, given raw text input. Intended for portfolio, educational, and demonstration purposes, and as a baseline fake-news detection component.

Downstream Use

Could be integrated into a larger content-moderation or fact-checking pipeline as one signal among several, rather than a standalone decision-maker.

Out-of-Scope Use

Not intended for production moderation decisions, legal/policy enforcement, or as the sole basis for judging content authenticity. Not evaluated on non-English text, satire, or news domains/formats outside the WELFake dataset.

Bias, Risks, and Limitations

• Trained on a single dataset (WELFake); may not generalize well to other fake-news domains, sources, or writing styles.
• Deduplication reduced train/test leakage substantially, but residual near-duplicate or topically similar content may still inflate reported metrics slightly.
• May reflect labeling biases or source-selection biases present in the original WELFake dataset.

Recommendations

Users should treat predictions as one signal, not ground truth, and independently verify claims rather than relying solely on this model's output — especially in any real-world or high-stakes deployment.

How to Get Started with the Model

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

tokenizer = AutoTokenizer.from_pretrained("GhostFaith/distilbert-fakenews")
model = AutoModelForSequenceClassification.from_pretrained("GhostFaith/distilbert-fakenews")

text = "Your news article text here"
inputs = tokenizer(text, truncation=True, padding=True, max_length=256, return_tensors="pt")
with torch.no_grad():
    logits = model(**inputs).logits
pred = torch.argmax(logits, dim=-1).item()
print("Fake" if pred == 1 else "Real")

Training Details

Training Data

WELFake news dataset (~71,537 rows total). Deduplicated on article text (~9,337 duplicate rows removed) before splitting, to prevent identical or near-identical articles from appearing in both train and test sets. Split 80/10/10 into train/validation/test.

Training Procedure

Fine-tuned using Hugging Face Trainer.

Preprocessing

Text tokenized with the distilbert-base-uncased tokenizer, truncation enabled, padded to a fixed max length.

Training Hyperparameters

• Epochs: 2
• Train batch size: 8 (per device)
• Eval batch size: 16 (per device)
• Warmup steps: 500
• Weight decay: 0.01
• Training regime: fp16 mixed precision

Speeds, Sizes, Times

Trained on a single Colab T4 GPU; full training run (2 epochs, ~57K training examples after split) completed in well under an hour.

Evaluation

Testing Data, Factors & Metrics

Testing Data

Held-out test split (10%) from the deduplicated WELFake dataset — not seen during training or validation.

Metrics

Accuracy and F1 score, plus a full precision/recall/F1 breakdown per class (real vs. fake).

Results

Confusion matrix (rows = actual, columns = predicted, order = [Real, Fake]):

[[3453   28]
 [  27 2712]]

Summary

The model performs strongly on held-out, deduplicated test data, with balanced errors across both classes (28 false positives, 27 false negatives out of 6,220 test examples) — no systematic bias toward over- or under-flagging either class.

Environmental Impact

• Hardware Type: NVIDIA T4 GPU (Google Colab)
• Hours used: < 1 hour
• Cloud Provider: Google Colab
• Compute Region: Not tracked
• Carbon Emitted: Not estimated

Technical Specifications

Model Architecture and Objective

DistilBERT (6-layer transformer encoder, distilled from BERT-base) with a sequence classification head (pre-classifier + classifier layers) for binary text classification.

Compute Infrastructure

Hardware

Single NVIDIA T4 GPU (Google Colab free tier).

Software

Python, Hugging Face transformers, datasets, PyTorch.

Model Card Contact

See the GitHub repository for contact information.