Datasets:

torux
/

EditCLEVR

language: en
license: cc-by-4.0
task_categories:
  - image-to-image
tags:
  - object-centric-learning
  - intervention-faithfulness
  - CLEVR
  - synthetic
  - computer-vision
pretty_name: EditCLEVR Phase 1
size_categories:
  - 10K<n<100K

EditCLEVR Phase 1

EditCLEVR is a synthetic benchmark for evaluating intervention faithfulness in object-centric representations. Each example is a paired before/after scene where exactly one object-level factor may change.

Dataset summary

~20k paired edits across six evaluation splits
Suites: atomic single-factor edits, no-edit controls, hard distractors, and CoGenT-OOD combinations
Per-object factors: color, material, size, shape
Artifacts per pair: before/after RGB images, instance masks (.npz), scene JSON, object attributes, edit metadata, and difficulty tags

Splits

Split	Purpose
`train`	Probe training
`val`	Validation
`test_id`	In-distribution atomic edits
`test_noop`	No-edit control pairs
`test_hard`	Hard distractor edits
`test_cogent`	CoGenT-OOD edits

Download

The dataset ships as a small set of .tar.gz archives (one per suite plus a splits bundle). The helpers below download and extract them into the original directory layout automatically.

git clone https://github.com/torux-bughunter/EditCLEVR.git
cd EditCLEVR
pip install -e ".[hub]"
python -m editclevr.download

Or from Python:

from editclevr.download import setup_dataset

setup_dataset()

Evaluate

pip install -e .
python3 -m editclevr.evaluation.run_evaluation

File layout

splits.json
atomic_id/
no_edit/
hard_distractor/
cogent_ood/
validation_report.json
phase1_manifest.json

splits.json stores relative paths to images, masks, and scene JSON files so the dataset can be moved across machines.

License

Dataset: CC-BY-4.0
Code: MIT

Citation

If you use this dataset, please cite:

Anuraag Gadehothur Karnam and Tarunesh Sathish. EditCLEVR: A Paired-Scene Intervention Benchmark for Compositional Faithfulness of Object-Centric Representations. ICML 2026 Workshop on Combining Theory and Benchmarks: Towards A Virtuous Cycle to Understand and Guarantee Foundation Model Performance, 2026. Paper URL work in progress.

See CITATION.cff in the code repository.