Datasets:

williamhe0712
/

CARScenes

license: cc-by-4.0
pretty_name: CARScenes
tags:
  - autonomous-driving
  - vision-language
  - scene-understanding
  - benchmark
  - croissant
task_categories:
  - image-classification
  - visual-question-answering
language:
  - en
size_categories:
  - 1K<n<10K
configs:
  - config_name: default
    default: true
    data_files:
      - split: train
        path: release/carscenes-v1/splits/train.jsonl
      - split: validation
        path: release/carscenes-v1/splits/dev-500.jsonl
      - split: test
        path: release/carscenes-v1/splits/gold-test-100.jsonl
  - config_name: full_release
    data_files:
      - split: train
        path: release/carscenes-v1/data/carscenes_v1_records.jsonl
  - config_name: agreement_subset
    data_files:
      - split: test
        path: release/carscenes-v1/splits/gold-agreement-25.jsonl

CARScenes

CARScenes is an annotation-only dataset and benchmark for structured scene understanding and failure analysis in autonomous driving.

Dataset Summary

Version: carscenes-v1
Records: 5,192
Sources: Argoverse1, Cityscapes, KITTI, nuScenes
Release type: annotations, schema, split manifests, benchmark tooling, Croissant metadata

Public Links

GitHub: https://github.com/Croquembouche/CARScenes
Paper PDF: https://github.com/Croquembouche/CARScenes/tree/main/paper/output

What Is Included

Canonical JSONL annotations
Machine-readable schema
Fixed split manifests
Gold benchmark split and agreement subset manifests
Croissant metadata
Evaluation and validation scripts

What Is Not Included

Upstream source images are not redistributed
Users must obtain the original images under the original dataset licenses and terms

Source Pool Scope

The released universe is the pre-annotation candidate pools stored in the CARScenes release artifacts, not the full upstream corpora.
Those candidate pools were created from downsampled / decimated front-camera subsets before annotation.
They are not exhaustive copies of the source datasets and are not intended to be statistically representative samples.
The exact historical decimation recipe is not claimed as part of the benchmark specification.

Splits

train: 4,592 records
dev-500: 500 records
gold-test-100: 100 records
gold-agreement-25: 25 records

License

CARScenes annotations, schema, split manifests, and metadata are released under CC BY 4.0.

This license does not cover source images from Argoverse1, Cityscapes, KITTI, or nuScenes, which remain under their original licenses.

Citation

@misc{he2026carscenes,
  title={CARScenes: A Benchmark for Severity-Aware Scene Understanding in Autonomous Driving},
  author={Yuankai He and Weisong Shi},
  year={2026},
  url={https://huggingface.co/datasets/williamhe0712/CARScenes}
}

Benchmark Use

Use dev-500 for prompt selection or pilot calibration.
Use gold-test-100 for headline model reporting.
Use gold-agreement-25 only for independent second-pass agreement analysis.