WORLD MODEL · DATA & BENCHMARK LAB

Data & Benchmark Lab

Explore practical dataset patterns, split strategies and benchmark designs for predictive, action-conditioned and long-horizon world models.

Dataset & Benchmark Patterns

Best for

Main risk

Required fields / signals

Optional fields / signals

Benchmark stack

PredictionMeasure immediate next-state quality.
RolloutMeasure error growth over several horizons.
Action fidelityVerify that actions produce correct future changes.
GeneralizationUse held-out environments, tasks, objects or embodiments.
Planning utilityMeasure whether the model improves decisions.
EfficiencyTrack latency, memory and rollout throughput.

Included practical files

transition_schema.json · episode_schema.json · benchmark_config.json · evaluation_template.csv

World Model · Cooperation, research and ecosystem partnerships: agenten@magenta.de