Generate video ad copy grounded in product facts and evaluate it with verifiable measures of required-fact retention and unsupported claims.

Test whether numerical claims in social ads are genuinely supported by citations with explicit metric definitions and conditions.

Extract source-grounded developments and timelines from international news reports while keeping cross-report differences explicit.

The street dance action estimation video dataset is dedicated to improving the accuracy of dance action recognition and estimation.

Ground professional music licensing Q&A and key clause analysis in the contract text.

Train models to predict the next MOBA lane state from synthetic events and structured pre-event conditions.

Turn out-of-home advertising compliance requirements into structured constraints for precise extraction of brand rules, disclosures, and placement restrictions.

Train models to answer breaking-news questions with traceable evidence by following how reports confirm, revise, or qualify facts.

Train models to spot outdated or inapplicable social media ad claims with structured temporal, geographic, availability, and evidence checks.

Train models to trace breaking news across reports and reconstruct evidence-backed event timelines.

Trace out-of-home ad claims from claim extraction through evidence linking to an overall fact-review conclusion.

Train models to turn requirements for filter appearance, scenes, and interactions into JavaScript code for AR social filters.

Follow a brand creative agent from review feedback through multi-tool iteration to revised concepts and recommendations.

Train code models on practical mobile shooter scenarios involving touch controls, aiming, and firing interactions.

Extract structured events from international news and trace key attributes back to evidence in the source text.

Link news claims to their sources and supporting evidence so models can preserve reporting differences without deciding which account is true.

Train models to predict hero health, shield, and control-state changes from combat events and rules with structured examples for MOBA state modeling.

Help models detect outdated citations and evidence by testing whether social ad claims align with their sources over time.

Train on practical Lua and Luau completion examples for core ad and in-app purchase flows in mobile casual games.

This dataset focuses on speech recognition tasks for academic lecture scenarios.