FAVA Datasets

FAVA datasets include: annotation data and training data.

Dataset Details

Annotation Data

The annotation dataset includes 460 annotated passages identifying and editing errors using our hallucination taxonomy. This dataset was used for the fine-grained error detection task, using the annotated passages as the gold passages.

Training Data

The training data includes 35k training instances of erroneous input and corrected output pairs using our synthetic data generation pipeline.