Feb 13, 2026 Research
SkillRater: Untangling Capabilities in Multimodal Data
Which skills does your data teach?
SkillRater scores multimodal training data by the capabilities each example actually exercises — grounding, counting, reading, reasoning — so curation stops being guesswork.
Data mixes chosen by SkillRater train stronger models with fewer examples.
Capability labels at scale
A judge model tags millions of examples with fine-grained skills, turning an opaque pile of data into a priced menu of capabilities.
Less data, better models
Removing low-signal examples while keeping skill coverage improves benchmark scores and cuts training cost at the same time.
Open methodology
We publish the skill taxonomy and scoring prompts so the community can rate and trade datasets on capability, not vibes.