Liar Dataset

LIAR Dataset is a benchmark for fake news and political statement classification, containing 12,836 short statements labeled with six fine-grained truthfulness ratings.

## What Is the LIAR Dataset?

• Size: 12,836 labeled statements
• Source: PolitiFact fact-checking verdicts
• Labels: Six levels (pants-fire to true)
• Metadata: Speaker, party, context, history

## Why LIAR Matters

Fighting misinformation requires nuanced truth assessment—not binary true/false. LIAR provides fine-grained labels and real political context.

LIAR Label Distribution:
pants-fire ████████░░░░░░░░░ (extreme false)
false      ██████████████░░░
barely-true███████████░░░░░░
half-true  ██████████████░░░
mostly-true████████████░░░░░
true       █████████░░░░░░░░

Example Statement:
"Exposed DNC emails show Clinton camp rigged 
primary against Bernie Sanders."
Label: half-true (some truth, misleading framing)

Model Performance:
| Model | Accuracy |
|-------|----------|
| Human (agreement) | ~75% |
| BERT fine-tuned | ~27% (6-class) |
| With metadata | ~41% |
| Binary (true/false) | ~60% |

Six-class classification is very challenging; most work focuses on binary.