What is World Model Data?
Environment interaction data for training world models, emphasizing temporal consistency, physical plausibility, and predictability.
World model data is long-horizon environment interaction: sequences long enough and consistent enough that a model can learn how the environment evolves. The unit is a continuous sequence rather than a clip, and the length is the specification.
What distinguishes it from video data is causal consistency. A world model has to predict what happens next given what happened before, which means the sequence cannot contain cuts, jumps or discontinuities that break the causal chain.
The uses are prediction and planning: models that forecast environment state, and agents that plan inside a learned model rather than in the real environment. Both depend on sequences long enough to contain an outcome.
What buyers get wrong
World models need long-horizon, causally consistent data; short fragmented clips are worthless. Collection design has to work backward from the required sequence length.
The specification detail that decides everything
The duration and sampling rate of a single sequence are the core specification for this data.
This is the line item that most often separates a dataset that works from one that gets re-ordered. It belongs in the specification before production starts, not in the delivery review.
How we source it
We do not hold inventory in this category. A requirement comes in, we match it to a producer who can meet the specification, and we agree the quality bar with both sides before recording begins. A pilot batch goes first so misalignments surface early.
Scenarios this covers
Buyers rarely ask for a category by name. They describe a situation, and the situation is what tells us whether this category is the right one. If yours is in the list, name it in the brief and it goes straight into the specification. If it is not, describe it anyway, because it usually means the requirement sits in a neighbouring category.
What good World Model Data looks like
World model data is judged on sequence length and causal integrity, in that order.
-
Sequence duration stated, and long enough to contain outcomes
A clip that ends before the consequence of an action appears cannot train prediction.
-
Sampling rate stated and constant
A variable rate breaks the temporal structure the model is learning, and the rate has to be documented per sequence.
-
No cuts or discontinuities
The sequence is continuous, and any unavoidable break is marked rather than smoothed over.
-
Action and state recorded together
What the agent did and what changed, because prediction without the action is a different task.
-
Physical plausibility verified
Sequences that violate the environment's own rules are removed or flagged, since a model learns the violations as readily as the rules.
Public datasets in this category
These are openly documented resources, not products we sell. Each one covers part of the category, and knowing where each stops is usually what decides whether a project needs a collection of its own.
-
Something-Something
Short clips of humans performing simple object interactions, annotated for the action. It is used for physical-reasoning work because the actions are simple and the labels are about what changed.
-
Kinetics
A large video action-recognition corpus spanning hundreds of human action classes. Broad, and it is about recognising actions rather than about predicting physical outcomes.
-
Robot simulation suites
Physics simulators with bundled scene and task sets, used to generate interaction data at scale. The data is generated rather than recorded, and the physics is an approximation of the real world.
-
Why this category is still forming
World models are a recent research direction and there is no settled corpus design for them. What exists is either action recognition or simulation, and neither was built for predicting what happens next.
Questions to ask a supplier
Questions that establish whether the sequences are long enough and continuous enough to be useful.
- What is the duration of a single sequence, and what is the shortest one delivered?
- At what rate is the environment sampled, and is it constant across a sequence?
- Are there cuts or discontinuities, and how are they marked?
- Is the agent's action recorded alongside the state?
- How was physical plausibility checked?
The vague answer in this category concerns the outcome. A supplier who says the corpus records interactions without saying whether the resulting state was recorded has described an action recognition dataset under a different name. A world model learns what changes, so the outcome is the label that matters.
A verifiable answer states the recorded state variables, the physics parameters where known, and whether failed and partial outcomes were kept. Ask for the share of episodes that did not succeed, because a corpus of clean completions teaches an optimistic physics that fails on contact.
Technical attributes
The items below are the ones a specification can hold a producer to, and the ones that decide whether a delivery is usable for training rather than only present.
| Sequence duration | Stated, with the minimum as well as the typical |
|---|---|
| Sampling rate | Constant, documented per sequence |
| Continuity | Unbroken, with any break marked |
| Action channel | Recorded alongside state |
| Physical plausibility | Verified, with the method stated |
| Delivery format | Continuous sequences with a per-sequence manifest |
The outcome variables column decides whether the corpus can train prediction at all, and it is the column most often reduced to a success flag. A binary outcome says whether something worked, not what changed. Position, orientation and contact state are what a predictive model needs, and they have to be captured during the episode.
The values come from the physical system being modelled. A rigid object on a table and a deformable object in a gripper have different state spaces, and the specification has to name which one. A supplier proposing a generic state list is proposing a corpus that will need re-collection for any specific use.
A mistake here is not repairable from the delivered files. If the intermediate state was not recorded, no amount of processing recovers it, and the corpus can only support the recognition tasks it was not bought for. This is the category where the recording design is the entire product.
How World Model Data is collected
Collection design works backward from the sequence length the model needs.
- Sequence length determined first. The required horizon is derived from what the model has to predict, and the collection is designed to produce sequences at least that long.
- Environment and dynamics selection. The environment is chosen for the dynamics the model has to learn, and its rules are documented so plausibility can be checked against them.
- Continuous capture without interruption. Recording runs without cuts, and where an interruption is unavoidable the break is recorded rather than hidden.
- Action logging in parallel. The agent's actions are logged on the same clock as the observations, because prediction conditioned on action is the task.
- Plausibility review and delivery. Sequences are reviewed against the environment's rules, violations are flagged or removed, and the review method is documented.
After collection, the state trajectories are checked for sensor drift, failed episodes are retained and labelled, and the files are packaged with the physics parameters that were measured. The drift check is what keeps the predicted outcome from diverging from the recorded one.
The pilot for world model data is a small set of episodes recorded with the full state capture running, because the property being tested is whether the outcome variables can be recorded at all without disturbing the task. A short run shows the sensor drift and whether the intermediate states are actually captured. Recording the actions without the outcomes produces a corpus that looks complete and supports only recognition.
Acceptance is checked on the completeness of the state trajectories and on the retained failed episodes, verified before the batch is accepted. The specification should name the state variables and the physics parameters before collection starts. Where the pilot shows a variable cannot be measured on the platform, the task is re-scoped rather than the variable dropped, since a missing outcome cannot be recovered afterwards.
How a delivery goes wrong
The defects below are the ones that recur across this category. Each is decided at collection or annotation time, which is why none of them can be repaired in delivery review.
-
Outcome was not recorded
A world model learns what changes. A corpus that records only the action and the scene, without the resulting state, cannot train prediction and is really an action-recognition dataset under a different name.
-
No counterfactual or failure examples
Predicting what happens when an action succeeds is easier than predicting what happens when it fails. Without failed and partial outcomes the model learns an optimistic physics.
-
Physics parameters are unknown
Mass, friction and compliance determine the outcome. Where the corpus does not document them, the model cannot generalise across objects, and the gap is invisible in aggregate metrics.
Why this matters to buyers
Prediction and planning are where the value is. A model that forecasts environment state lets an agent evaluate a plan without executing it, which is the difference between learning by trial in the real world and learning inside a model.
The second driver is that short clips do not contain the signal. A three-second clip shows a state; a thirty-second sequence shows a consequence. The learning target is the consequence, which is why sequence length is the core specification rather than a delivery detail.
A predictive model trained without outcomes learns to describe rather than to predict. It performs well on the recognition benchmark and fails whenever it is asked what happens next, which is the entire purpose. The corpus cannot be extended to fix this, because the missing variables were never recorded.
The second cost is research time. Teams working in this category are usually building something new, and a corpus that turns out to support a different task costs a cycle of experiments before the mismatch is accepted. That cycle is longer than the collection would have been.
A corpus with complete state trajectories gives a model that predicts what happens next rather than only describing what happened, which is the entire purpose in this category. The retained failures mean the predicted physics includes contact and partial success rather than only clean completions.
The second gain is that the recording design is documented. The variables captured and the physics measured are the product in this category, and a delivery that records them can be extended to a second task or platform without repeating the design work. That record is what makes the corpus a platform rather than a one-off.
Specification at a glance
| Category | World Model Data |
|---|---|
| Group | New data paradigm |
| Critical spec | The duration and sampling rate of a single sequence are the core specification for this data. |
| Included by default | Pilot batch, metadata schema, source and consent documentation |
| Pricing | Quoted from specification. No published rates |
Related categories
-
Embodied AI Data
Multimodal data generated as robots and embodied agents operate in real environments: vision, action trajectories, and language instructions.
-
Synthetic Data
Data generated by models or simulation engines, used to fill gaps where real data is scarce, cover long-tail scenarios, and control cost.
Questions about World Model Data
Why is sequence length the core specification?
Because the learning target is what happens next, and a sequence that ends before the consequence appears does not contain it. Everything else about the corpus is secondary to whether the sequences are long enough to include an outcome.
Can we assemble long sequences from short clips?
No. Concatenating clips creates discontinuities at the joins, and a model learns the joins as readily as the dynamics. Continuity has to come from the capture.
What does physical plausibility checking involve?
Reviewing each sequence against the environment's documented rules, looking for states that could not have arisen from the preceding ones. It is done because a model learns violations as readily as regularities.
Does sampling rate matter as much as duration?
It matters differently. Duration determines whether an outcome is present; rate determines whether the dynamics are observable. A variable rate is worse than a low one, because the model cannot infer the time step.
Do you have World Model Data available now?
No. We do not hold inventory. Everything is produced against a specification. That means the timeline starts when the spec is agreed rather than immediately, and it also means the data matches your requirement instead of approximately matching something already built.
Can this be combined with other categories in one delivery?
Yes, and it is common. A single specification can cover several categories across several languages, with one pilot batch and one quality bar. It usually reduces total cost because contributor recruitment and project setup are shared.
What does the consent documentation cover?
Signed contributor consent stating the intended use, the collection methodology, and the annotation guideline the data was produced under. For projects involving EU data subjects the chain is built to satisfy GDPR, including the transfer mechanism. The terms for moving the data across a border.
Request World Model Data
Tell us the language, the hours, and what the data needs to look like. You will get a real number and a real timeline, not a range. If we cannot source it well, we will tell you that instead.
- Pilot batch before the full run, so problems surface early.
- Consent documentation delivered with the data.
- No medical or clinical data. No recorded telephone calls.
- You send the spec Language, hours, recording conditions, and the deadline.
- We reply within two business days A real number and a real timeline, not a range.
- You approve a pilot batch A small run first, so problems surface before the full order.
- Full run, then delivery with documents Audio, transcripts, and the consent record for every speaker.