What it is
Large MT hallucination benchmark with span-level annotations across 38 language pairs for training and evaluating detection models. 350,959 professionally annotated MT sentence pairs across 38 language pairs, each with binary hallucination labels, span annotations, and hallucinated-character scores in train/dev/test splits.
Key results
- 350,959 annotated samples
- 38 language pairs with train/validation/test splits
- CC BY-NC-ND 4.0; parquet text/tabular data
Useful for training and benchmarking models that detect hallucinated content in machine translation outputs.