Evaluate how reliably models verify industrial ventilation specifications against product-page evidence.

Use context-rich counseling excerpts and intent annotations to help models understand patients' needs and expressions during health behavior change.

Evaluate prescription safety review with structured pediatric dental cases and labeled medication constraint violations.

Train assistants to understand hot pot restaurant requests with chat intents spanning reservations, ordering, dining service, and checkout.

Reconstruct incident cause-and-effect chains from logs to help models identify ITSM root causes more accurately.

Evaluate evidence weighting, confidence calibration, and cautious scientific reasoning in challenging simulated underwater conditions.

Connect GRE passages with external web evidence and information-boundary analysis to help models explain reading context with verifiable support.

Train models to verify balance adjustments and reconciliation results using cash ledgers, bank statements, and outstanding items.

Train models to answer French vocabulary questions with clear sense distinctions, natural collocations, and example-based evidence.

Train models to turn application materials and policy sources into complete, evidence-traceable reviews of personal loan decline decisions.

Evaluate how well models interpret numerical conditions, applicable configurations, and restrictions in scaffolding design and load specification texts.

Evaluate whether models can clarify question meaning and provide well-grounded answers while maintaining context across question bank dialogues.

Train models to trace code dependencies, control flow, and delivery outcomes when notifications fail and retry or switch channels.

Train models to analyze co-simulation scheduling and synchronization logic with focused, structured code understanding samples.

Train models to determine whether health and wellness advertising claims are genuinely supported by their cited evidence, including cases of weak support and overgeneralization.

Train models to acknowledge evidence gaps, avoid fabrication, and ask useful follow-up questions with multi-turn examples built around insufficient evidence.

Train models to track active conditions and respond accordingly with multi-turn examples covering persistent, updated, and invalidated rules.

Step-level constraint labels help models catch workflow violations from order intake through serving.

Connect reactor actions, process trajectories, and product metrics to support simulation-based research on selectivity optimization.

A high-quality text conversation dataset containing hotel ratings, reviews, and replies.