---
language: en
tags:
  - fake-news
  - text-classification
  - distilbert
  - pytorch
license: mit
datasets:
  - clmentbisaillon/fake-and-real-news-dataset
metrics:
  - f1
  - accuracy
  - roc_auc
---

# FakeGuard — Fake News Detector

Fine-tuned DistilBERT for binary fake news classification.

**CMPE 258 — Deep Learning · San José State University · Spring 2026**

## Model Details

| Property | Value |
|----------|-------|
| Base model | distilbert-base-uncased |
| Task | Binary text classification (Fake/Real) |
| Training samples | 2,749 (balanced subsample of 44K articles) |
| Epochs | 2 |
| Max sequence length | 256 |
| Dropout | 0.3 |
| Optimizer | AdamW (lr=3e-5, wd=0.01) |
| LR schedule | Cosine annealing |

## Performance

| Metric | Value |
|--------|-------|
| Test Accuracy | **0.9933** |
| Test Macro F1 | **0.9933** |
| Test ROC AUC | **0.9988** |
| Cohen's Kappa | **0.9867** |

## Dataset

Kaggle "Fake and Real News Dataset" (Emine Yetm / clmentbisaillon).

**Critical fix applied:** Removed `(Reuters)` dateline prefix from real articles before training to prevent data leakage.

## Live Demo

Try it at: https://huggingface.co/spaces/gpreetam236/DL_Final

## Architecture

```
DistilBERT Encoder (6 layers, 768-d) → [CLS] token
→ Dropout(0.3) → Linear(768→256) → LayerNorm → GELU
→ Dropout(0.3) → Linear(256→2) → Logits
```

## W&B Experiment Dashboard

https://wandb.ai/gowripreetham23-san-jose-state-university/fakeguard