Datasets:
File size: 2,735 Bytes
6458d51 f691d6f 6458d51 f691d6f 6458d51 f691d6f 6458d51 f691d6f 6458d51 ce817b4 6458d51 3f8ef67 6458d51 3f8ef67 6458d51 3f8ef67 6458d51 3f8ef67 6458d51 3f8ef67 6458d51 734efa9 f691d6f | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 | ---
language: en
license: cc-by-4.0
size_categories:
- 10K<n<100K
task_categories:
- image-to-image
pretty_name: EditCLEVR Phase 1
tags:
- object-centric-learning
- intervention-faithfulness
- CLEVR
- synthetic
- computer-vision
---
# 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.
Paper: [EditCLEVR: A Paired-Scene Intervention Benchmark for Compositional Faithfulness of Object-Centric Representations](https://huggingface.co/papers/2607.22705)
## 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.
```bash
git clone https://github.com/torux-bughunter/EditCLEVR.git
cd EditCLEVR
pip install -e ".[hub]"
python -m editclevr.download
```
Or from Python:
```python
from editclevr.download import setup_dataset
setup_dataset()
```
## Evaluate
```bash
pip install -e .
python3 -m editclevr.evaluation.run_evaluation
```
## File layout
```text
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](https://creativecommons.org/licenses/by/4.0/)
- **Code:** [MIT](https://github.com/torux-bughunter/EditCLEVR/blob/main/LICENSE)
## 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`](https://github.com/torux-bughunter/EditCLEVR/blob/main/CITATION.cff) in the code repository. |