Benchmarks & research

JEV Benchmark – CLASH contradiction detection

Runs the CLASH cross-modal contradiction test against Jev with each image replaced by its COCO caption, then reports accuracy and modality bias.

What it does

Because Jev takes no images, a task designed to be cross-modal becomes text against text — worth knowing before reading the score.

Maker-reported (not independently measured by JevMade): 1,289 human-verified test samples · 859 samples (alt_caption runs)

Repository ↗

Source checked 2026-09-24 — opened the GitHub repository directly.