SeerWorld-Data is a research dataset for training generative video world models to represent world evolution beyond the current field of view. Each sample is a single-take video in which the camera observes a subject, moves away to reveal an event, and later revisits the original region after the subject has changed in response.

The dataset accompanies the paper “SeerWorld: Towards Out-of-Sight World Evolution in Generative Video Models.” SeerWorld-Data is released separately from SeerWorld-Bench because the two resources serve different purposes and use different schemas: SeerWorld-Data provides training supervision, while SeerWorld-Bench provides a controlled evaluation protocol.

Dataset summary

Split	Videos
Train	11,817
Additional	301
Total	12,118

The release contains 6,413 ten-second videos with three segment prompts and 5,705 fifteen-second videos with four segment prompts. All released videos are represented at 24 fps in the metadata.

Repository structure

SeerWorld-Data/
├── data/
│   ├── train/
│   │   ├── metadata.jsonl
│   │   └── videos/000/ ... videos/011/
│   └── additional/
│       ├── metadata.jsonl
│       └── videos/011/ ... videos/012/
├── training_cache/ltx-2.5/      # optional precomputed inputs
├── THIRD_PARTY_NOTICES.md
└── LICENSE

Videos are grouped into 1,000-ID buckets to keep Hub directories manageable. Each file_name is relative to the metadata.jsonl file in the same split.

Metadata schema

Each line in metadata.jsonl is a JSON object with the following fields:

Field	Description
`id`	Stable public sample identifier.
`file_name`	Relative path to the MP4 video.
`split`	`train` or `additional`.
`duration_s`	Nominal generation duration in seconds: 10.0 or 15.0 (not an ffprobe measurement).
`fps`	Video frame rate used by the release metadata.
`segment_frames`	Temporal window length used for segment-level training supervision.
`segment_prompts`	Ordered prompts describing consecutive video segments.
`global_prompt`	A compact description of the complete video trajectory.
`subject_type`	Broad subject class: `human`, `animal`, or `game_character`.
`event_type`	Broad event class: `environmental_hazard`, `animal_threat`, `human_activity`, or `game_threat`.
`reaction_type`	Scenario-level reaction category; see vocabulary below.

The three type fields come from the original structured generation specifications, not independent labels verified against every rendered video. An intended event or reaction may be incompletely realized. They are useful for filtering training scenarios, but should not be treated as ground-truth evidence of successful execution. Subject classes distinguish real-world human/animal scenarios from game-character scenarios; game_character is a domain grouping, not a biological category.

Reaction vocabulary: flee, hide, take_flight, retreat, freeze_then_flee, move_aside, swim_away, climb, retreat_to_entrance, take_cover, and defensive_posture. These are the existing 11 reaction categories, not fine-grained action descriptions. In particular, the legacy ALERT label means moving to one side in the generation taxonomy and is exported as move_aside.

Example:

{
  "id": "seerworld_000000",
  "file_name": "videos/000/seerworld_000000.mp4",
  "split": "train",
  "duration_s": 10.0,
  "fps": 24,
  "segment_frames": 81,
  "segment_prompts": ["...", "...", "..."],
  "global_prompt": "...",
  "subject_type": "human",
  "event_type": "environmental_hazard",
  "reaction_type": "take_cover"
}

Loading

from datasets import load_dataset

dataset = load_dataset(
    "videofolder",
    data_dir="hf://datasets/Tommy-Bie/SeerWorld-Data/data",
)
train_set = dataset["train"]
additional_set = dataset["additional"]

The raw files can also be downloaded directly and consumed using the per-split JSONL metadata. The 301 additional samples were not used to train or validate the paper model. They are provided as extra valid training cases and may be merged into the training split for downstream use.

Optional LTX-2.5 training cache

The repository also provides precomputed inputs for reproducing the released SeerWorld training recipes without re-encoding every video:

Component	Contents	Files	Approx. unpacked size
`core`	Full-video latents and entity-reference latents	24,236	53 GiB
`conditions`	Per-segment text features	42,059	494 GiB

The two components are independent downloads. The condition cache is optional; users may instead run the text encoder locally. Global-prompt condition files are intentionally omitted because the released recipes use only per-segment conditions. See training_cache/ltx-2.5/README.md for checksums, extraction, paths, and tensor details.

Intended use

SeerWorld-Data is intended for research on video generation, video world models, segment-autoregressive generation, long-horizon consistency, event-conditioned generation, and out-of-sight world evolution. Eligibility is determined by use: academic, nonprofit, individual, and company researchers may use it for non-commercial research, but it may not be used to train, fine-tune, evaluate, improve, or deploy a commercial model.

Data construction and quality control

The videos were constructed with generative AI under structured scenario specifications. Samples were screened for scene structure, camera motion, event realization, revisit behavior, visual integrity, and harmful content. Segment-level and global captions were produced from the rendered videos and checked as part of the data pipeline.

Automated and model-assisted screening cannot guarantee that every sample is artifact-free. Some videos may contain generation artifacts, imperfect event rendering, identity drift, or captions that describe an event near a segment boundary. These long-tail imperfections should be considered when training or analyzing models.

Limitations

The current release focuses on a single event followed by a single revisited response.
Camera trajectories primarily use horizontal move-away and move-back motion, a common revisit pattern, rather than arbitrary camera paths.
The dataset contains generated visual content and should not be treated as a substitute for real-world behavioral evidence.
Captions describe the rendered content and do not constitute factual claims about real people or events.

License

SeerWorld-Data is released under the custom SeerWorld Research-Only License v1.0 (research-only-v1.0). The dataset, its precomputed cache, and models trained from it may be used only for non-commercial research. Commercial-model training, evaluation, improvement, and deployment are prohibited. See LICENSE for the complete terms.

The optional cache does not include model weights. Users must obtain required upstream models separately and comply with their terms. See THIRD_PARTY_NOTICES.md; the dataset license does not replace third-party licenses or grant rights the authors do not hold.

Datasets:

Tommy-Bie
/

SeerWorld-Data

You need to agree to share your contact information to access this dataset

SeerWorld-Data