README.md

---
license: openmdw-1.1
language:
  - en
tags:
  - video
  - autonomous-driving
  - simulation
  - multiview
configs:
  - config_name: default
    drop_labels: true
    data_files:
      - split: test
        path:
          - wsm/*/*.mp4
          - metadata.jsonl
---

# Cosmos3 Driving Simulation Evaluation

## Dataset Summary

This dataset contains **91 driving scenes** with **11 camera views per scene**:
1,001 World Scenario Map (WSM) control videos and 1,001 matching timestamped
caption JSON files.

## Dataset Description

The dataset provides structured scene controls and text prompts for multiview
driving generation and evaluation. The WSM controls are derived from real-world
driving data collected by NVIDIA and represent lane lines, object bounding
boxes, and traffic-light states.

**Included:** WSM 4.0.1 rendered control videos, per-camera captions, scene/view
manifests, measured video properties, and file checksums.

**Not included:** source RGB camera images or videos, LiDAR data, calibration,
raw scene annotations, or model-generated outputs. WSM videos are abstract
renderings of scene structure, not camera recordings. Original or restyled
first-frame images are not included.

The released manifest identifies no separate standard or long-rollout subsets;
control videos are 11.7–21.9 seconds long.

## Dataset Owner(s)

NVIDIA Corporation

## Dataset Creation Date

2026-10-20

## Versioning

V1.0

No previous version.

## License/Terms of Use

This dataset is released under the [OpenMDW License Agreement, version 1.1
(OpenMDW-1.1)](https://openmdw.ai/license/1-1/).

## Intended Usage

This dataset is intended for autonomous vehicle-related use cases only, including
both commercial and non-commercial use in accordance with the applicable license
terms. Its size and distribution across camera views support synthetic driving
scene evaluation.

Public text-to-video evaluation of Cosmos driving simulation using WSM controls
and accompanying text prompts. Specifically:

- evaluating Cosmos3 driving simulation conditioned on WSM controls.

This dataset is not intended for training perception or driving models.

## Dataset Characterization

**Data Collection Method:** Automatic/Sensors.

Sensor-collected driving data is processed into WSM controls.

**Labeling Method:** Automated.

Human reviewers provide quality assurance for the automated annotations. Processing steps:

1. NVIDIA in-house models annotate lane lines and object bounding boxes on the source footage.
2. Human reviewers check the annotations for quality.
3. A script converts the reviewed annotations into WSM controls.
4. An internal NVIDIA caption model generates a text caption per camera view, used as a text prompt for evaluation.
5. All camera imagery is discarded. Only the WSM controls and text prompts are released.

## Dataset Format

**Modality:** Video and Text.

WSM control videos in mp4 format, at 30fps. Each video shows the scene layout over time: lane lines, object bounding boxes, and traffic-light states.

Text captions are provided as 11 separate JSON files per scene, one per camera view.

## Dataset Quantification

Record count: 91 multiview sequences (11 camera views per sequence).

Size bin: `n<1K`, where each record is one sequence.

Video length: 351–657 frames (11.7–21.9 seconds) at 30fps.

Resolution:
- Seven standard cameras: 1924 × 1084 for 53 sequences, or 1920 × 1080 for 38 sequences.
- Four fisheye cameras: 1720 × 1080 for 53 sequences, or 1350 × 1080 for 38 sequences.

Total storage size: approximately 2.59 GB (decimal).

## Limitations

- **Small sample.** 91 scenes cover a limited range of road types, regions, traffic densities, and weather. Results on this evaluation subset may not generalize to other driving scenarios.
- **One conditioning format.** Results apply to WSM conditioning and may not transfer to other conditioning formats, such as depth, segmentation, or edge maps.
- **Annotation error.** Lane lines and bounding boxes come from automated models with human QA. Some missed or misplaced objects may remain.
- **Layout only.** WSM encodes scene geometry and traffic-light state. It does not encode appearance, lighting, weather, or road-surface detail.

## Reference(s)

- [WSM representation example](https://github.com/NVIDIA/cosmos/blob/main/assets/demos/transfer_worldscenario.gif).
  The top half shows WSM controls; the bottom half shows Cosmos-generated video. The released dataset contains WSM controls and text prompts, without generated video.

## Ethical Considerations

NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. Developers should work with their internal developer teams to ensure this dataset meets requirements for the relevant industry and use case and addresses unforeseen product misuse.

**Privacy:** The dataset contains no camera images, faces, license plates, or legible text. Road agents appear only as bounding boxes.

Please report quality, risk, security vulnerabilities or NVIDIA AI Concerns [here](https://www.nvidia.com/en-us/support/submit-security-vulnerability/).

## Files

```text
wsm/scene-001/camera_front_wide_120fov.mp4
captions/scene-001/camera_front_wide_120fov.json
...
manifest.jsonl
metadata.jsonl
checksums.json
```

- `manifest.jsonl`: one record per scene, grouping all 11 views by `camera_key`,
  with repository-relative `wsm_path` and `caption_path` values.
- `metadata.jsonl`: one record per camera view, linking its control video to the
  complete caption text and measured video properties for the dataset viewer.
- `checksums.json`: SHA-256 and byte size for every other release payload file.
- `scene-001` through `scene-091` are release-local identifiers.

The viewer has **1,001 rows**, not 1,001 independent scenes. Group rows by
`sample_id` to recover the 91 multiview scenes. The `test` split contains the
entire evaluation collection; no training split is supplied.

## Browse the dataset

Use the built-in **Dataset Viewer / Data Studio** on this dataset page, with
configuration `default` and split `test`. Each row contains one playable WSM
video and its matching timestamped caption.

- Search for a scene ID such as `scene-001` to find its 11 camera views.
- Use `camera_key` to identify the view, for example
  `camera_front_wide_120fov`.
- Read `caption` for the complete prompt. `caption_path` points to the matching
  structured JSON in the **Files** tab.
- Check `width`, `height`, `fps`, `num_frames`, and `duration_seconds` for the
  native control-video properties. `caption_start_frame` and
  `caption_end_frame` describe the caption window, not the full video duration.
- Download a complete scene using the command in the **Download** section below.

The built-in viewer displays camera videos individually; it is not a
synchronized surround player. It shows only WSM controls and captions, not RGB
recordings or LiDAR. Hosted browsing requires the data to be on the dataset's
default branch and processed by Hugging Face; files proposed only in a pull
request are not a verified live viewer. Restricted repositories require access.

## Camera names

Each scene contains these camera keys. Match files by camera key, not directory
listing order.

```text
camera_front_wide_120fov
camera_cross_left_120fov
camera_cross_right_120fov
camera_rear_left_70fov
camera_rear_right_70fov
camera_rear_tele_30fov
camera_front_tele_30fov
camera_front_fisheye_200fov
camera_left_fisheye_200fov
camera_rear_fisheye_200fov
camera_right_fisheye_200fov
```

## Captions and timing

Each JSON preserves the structured caption format:

```json
{
  "caption_structured": {
    "chunk_0_300": {
      "caption": "[0.0s-2.0s]: ...",
      "start_frame": 0,
      "end_frame": 300
    }
  }
}
```

This is a schema illustration, not an additional sample. Use each file's actual
chunk keys, frame bounds and timestamped text.

The captions are the timestamped Tier1 version, with a privacy cleanup of
identifying text. They are **not a new environmental prompt rewrite**. The
cleanup removes transcribed identifying values, including license-plate values,
while retaining generic descriptions such as a vehicle having a visible plate.
Timestamp strings, frame bounds and other JSON fields are preserved.

Control videos retain their original bytes, frame rate, dimensions and duration.
**A control video may be longer than its caption window.** Use the caption's
frame bounds when selecting a generation interval; do not assume that a caption
describes the entire MP4. `metadata.jsonl` records measured control dimensions,
frame count, frame rate and duration. The upload does not crop, retime, resize or
re-encode the controls.

Captions are machine-generated descriptions and may contain errors. Privacy
cleanup is not a correction of all scene descriptions or an independent
verification of physical camera calibration.

## Download

Download the files directly; there is no ZIP extraction step:

```bash
hf download nvidia/Cosmos3-Driving-Simulation-Evaluation \
  --repo-type dataset --local-dir driving-simulation-evaluation
```

For one scene, download its two folders plus the manifests:

```bash
hf download nvidia/Cosmos3-Driving-Simulation-Evaluation \
  --repo-type dataset --local-dir driving-simulation-evaluation \
  --include 'wsm/scene-001/*' 'captions/scene-001/*' \
            'manifest.jsonl' 'metadata.jsonl' 'README.md' 'checksums.json'
```

Authentication is required while repository access is restricted. Downloaded
metadata uses relative paths and contains no internal storage locations.