Research guide
A Practical World-Model Dataset Specification
The minimum episode, stream, timebase, provenance, rights, and quality fields needed for a portable trajectory release.
- For
- Dataset architects and research infrastructure teams
- Reviewed
- July 24, 2026
Direct answer
A portable specification makes every episode self-describing: what happened, which streams exist, how clocks align, where data came from, what rights apply, and which checks passed.
Make episodes first-class
The episode should identify task, environment, attempt boundaries, participants or agents, initial condition, terminal condition, and relationships to retries or alternate views.
Streams reference the episode timebase and declare modality, encoding, sampling behavior, units, coordinate frame, schema, and validity masks.
Preserve transformation history
Raw files, derived files, annotations, and normalized tensors should remain linked through checksums and transformation records. Each step names code version, parameters, inputs, outputs, and operator or automated system.
This lineage lets a buyer reproduce an export, trace a bad field, or withdraw affected derivatives without deleting unrelated material.
Make quality machine-readable
Quality is not one pass or fail flag. Record technical integrity, synchronization, semantic validity, rights eligibility, and release approval separately.
A failed task can be valuable training data. A completed task can still fail release because the action is hidden, state is missing, or rights are insufficient.
Ship documentation beside data
A release should include a JSON Schema, field dictionary, synthetic example, dataset card, citation file, Croissant metadata, loader checks, checksums, changelog, and known limitations.
ActionTrajectories publishes a starter specification for this purpose. It is a working interoperability baseline, not a claim that every environment shares one universal action space.