README.md

---
pretty_name: WORL LingBot Training Data
license: cc-by-4.0
language:
  - en
size_categories:
  - 1K<n<10K
tags:
  - worl
  - lingbot
  - world-model
  - image-to-video
  - camera-control
  - synthetic
configs:
  - config_name: default
    data_files:
      - split: train
        path: manifest.jsonl
---

# WORL LingBot Training Data

**WORL-Lingbot-Training-Data** is an OpenScene/revisit conditioning dataset for
training camera-controlled video world models. It contains text prompts,
initial scene images and camera trajectories in the portable WORL chain format.

Related model: [WORL-Lingbot](https://huggingface.co/Efficient-Large-Model/WORL-Lingbot).

## Dataset contents

| Item | Count / format |
| --- | --- |
| Training records | 2,218 |
| First-frame images | 1,419 PNG images and 799 JPEG images |
| Camera arrays | 4,436 NumPy files: poses and intrinsics |
| Chain lengths | 912 two-clip chains; 601 three-clip chains; 705 four-clip chains |
| Clip geometry | 81 frames per clip, stride 80, camera timeline at 16 FPS |
| First-frame dimensions | 1,419 images at 832 × 464; 799 images at 1536 × 1024 |
| License | CC BY 4.0 |

The package contains conditioning inputs, not generated video clips or
preference labels. The first-frame images retain their source dimensions;
the training pipeline is responsible for any required resizing and consistent
adjustment of camera intrinsics.

## Creation and scope

The WORL team created the OpenScene/revisit data and reports that the initial
scene imagery was produced using gpt-image. The portable bundle was prepared
from a 2,400-row source collection. It retains 2,218 records after excluding
182 rows whose source stride was incompatible with the LingBot geometry.

The source labels cover OpenScene, revisiting, rotation and shared-context
variants. This release provides the team's LingBot conditioning bundle; it is
not the full NVRM video preference corpus. The available model export does not
record an exact checkpoint-to-dataset binding.

## Download

```bash
hf download Efficient-Large-Model/WORL-Lingbot-Training-Data \
  --repo-type dataset \
  --local-dir data/WORL-Lingbot-Training-Data
```

Pin a Hub commit with `--revision` when reproducing an experiment.

## File layout

```text
manifest.jsonl          Training records in worl.chain.v1 format
assets/first_frames/    First-frame images
assets/camera/          Camera-to-world poses and camera intrinsics
assets.lock.json        Size and SHA-256 for every referenced asset
dataset-card.md         Short data card used by the WORL data preflight
PUBLICATION.json       Source snapshot, data identities and validation summary
RELEASE_MANIFEST.json   SHA-256 inventory of the public files
LICENSE.txt            Creative Commons Attribution 4.0 International
README.md              This dataset card
```

Each record contains a stable `id`, `sampling_index`, relative first-frame
path, camera metadata and ordered clip prompts with source frame windows.
`provenance.license` is `CC-BY-4.0`. Asset paths are relative to the dataset root.

| Camera field | Meaning |
| --- | --- |
| `coordinates` | `opencv_c2w`: camera-to-world convention |
| `poses_c2w` | Relative path to per-frame 4 × 4 pose matrices |
| `intrinsics` | Relative path to pixel-unit camera intrinsics |
| `fps` | Camera timeline frame rate, 16 FPS |

Example for reading one record locally:

```python
import json
from pathlib import Path
import numpy as np
from PIL import Image

root = Path("data/WORL-Lingbot-Training-Data")
with (root / "manifest.jsonl").open() as stream:
    record = json.loads(next(stream))

image = Image.open(root / record["first_frame"]["uri"]).convert("RGB")
camera = record["conditions"]["camera"]
poses = np.load(root / camera["poses_c2w"], allow_pickle=False)
intrinsics = np.load(root / camera["intrinsics"], allow_pickle=False)
prompts = [clip["prompt"] for clip in record["clips"]]
```

## Provenance and integrity

This public package derives from
[qyoo/lingbot-top2400-training-data at ee75eb9e711b](https://huggingface.co/datasets/qyoo/lingbot-top2400-training-data/tree/ee75eb9e711ba55ec9eac9356935a9dd21bf352f).
Images, camera arrays, prompts, record IDs, sampling order and clip windows are
preserved. Record provenance fields and documentation were updated for the
public release under CC BY 4.0.

The updated manifest and data card produce a new WORL dataset revision.
`PUBLICATION.json` records both the source identity and this release's identity;
the Hub commit is a separate version identifier. `assets.lock.json` covers all
6,654 media assets. The public metadata is checked with WORL's CPU dataset
preflight for schema, asset hashes, image geometry, camera arrays and sampling
buckets. This check does not represent a new training run.

## Intended use and limitations

The dataset supports research and development of camera-conditioned video
generation and world-model reinforcement learning. Synthetic imagery and
descriptions may contain visual errors, implausible geometry or bias. The
camera conditions do not establish physically accurate ground-truth scenes.
The bundle contains two- to four-clip chains and does not by itself reproduce
every experiment in the WORL manuscript.

## License and attribution

The dataset is licensed under
[Creative Commons Attribution 4.0 International](https://creativecommons.org/licenses/by/4.0/).
The full terms are in [LICENSE.txt](LICENSE.txt).

When sharing this dataset or an adaptation, credit the WORL team, provide the
dataset and license links, and indicate changes. A suggested attribution is:

> WORL team, *WORL LingBot Training Data*,
> https://huggingface.co/datasets/Efficient-Large-Model/WORL-Lingbot-Training-Data,
> licensed under CC BY 4.0.

Maintainer: Yuyang Zhao, NVIDIA Research Singapore & Efficient AI.