lie-detector-roberta

Lie Detector (RoBERTa)

This model is a fine-tuned version of roberta-base on the LIAR dataset, a benchmark for political fact-checking introduced in "Liar, Liar Pants on Fire" (Wang, 2017).

It classifies political statements into six categories: pants-fire, false, barely-true, half-true, mostly-true, true.

Alongside the statement, the model uses:

• Context and subjects
• Metadata: speaker, party, state (as embeddings)
• Numerical features**: historical counts of truthfulness

Results

• Test Accuracy (six-way classification): 40.5%
• Original paper accuracy: 27.4%

Example

label = lie_detector(
    statement="We’ve added more jobs than any time in history.",
    subjects="economy,jobs",
    speaker_name="Joe Biden",
    speaker_title="President",
    state="delaware",
    party_affiliation="democrat",
    history_barely_true=14,
    history_false=12,
    history_half_true=24,
    history_mostly_true=21,
    history_pants_fire=5,
    context_location="CNN Town Hall"
)