JevBench Multilingual
Decision models on five safety benchmarks in eight languages, each compared with its own English answers on the same cases.
Every language against English
Change from English, per model
points of accuracyEach cell compares a model with its own English answers on the same cases. Bold and underlined: the 95% paired bootstrap interval excludes zero. "…": the run is still going. Hover a cell for the numbers. Click a language to open it.
By benchmark
Accuracy at the 0.5 cut, mean over models. Colour runs per row, from the weakest language to the strongest.
Bar: accuracy in this language. Dashed line: the same model in English, on the same cases. Value: accuracy, then the change from English.
Per benchmark
Accuracy in this language. Hover a cell for English, the change, AUROC and how often the model said yes. A drop with an unchanged AUROC is a moved threshold, not lost skill.