LIAR
LIAR is a fake-news detection benchmark dataset introduced by Wang, comprising real political statements sourced from PolitiFact along with veracity labels. The dataset spans multiple sources and topics and is commonly used to evaluate the cross-domain generalization ability of fake-news detection models.
Timeline (last 90 days)
When all three models are transferred to the LIAR benchmark, they all drop to near-random ranking levels (ROC-AUC of only 0.54–0.57), with none able to beat the majority-class baseline.
The LIAR benchmark was published by Wang in 2017 and contains approximately 12,800 political statements from PolitiFact, annotated with six-level veracity labels.
State-of-the-art supervised models on the LIAR benchmark still hover around 28%–32% six-class accuracy.
All Facts (3)
When all three models are transferred to the LIAR benchmark, they all drop to near-random ranking levels (ROC-AUC of only 0.54–0.57), with none able to beat the majority-class baseline.
50%UnverifiedThe LIAR benchmark was published by Wang in 2017 and contains approximately 12,800 political statements from PolitiFact, annotated with six-level veracity labels.
50%UnverifiedState-of-the-art supervised models on the LIAR benchmark still hover around 28%–32% six-class accuracy.
50%