2.18.14
·
316 commits
to main
since this release
2.18.14 (2026-08-05)
Fix
- fix: mock zeroshot tasks must declare the text modality (#5093)
Zeroshot classification scores inputs against candidate text prompts, so
these tasks require a text tower. The audio, video and video+audio mocks
omitted "text" from metadata.modalities, so get_compatible_mock_tasks
selected them for audio-only and video-only models, which then fail with
KeyError on the input column. The image mock and every real zeroshot task
already declare text.
Co-authored-by: Hubert Lu <hubielu@email.com> (f7c910d)