Datasets:
Add EditCLEVR dataset card
Browse files
README.md
ADDED
|
@@ -0,0 +1,90 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
language: en
|
| 3 |
+
license: cc-by-4.0
|
| 4 |
+
task_categories:
|
| 5 |
+
- image-to-image
|
| 6 |
+
tags:
|
| 7 |
+
- object-centric-learning
|
| 8 |
+
- intervention-faithfulness
|
| 9 |
+
- CLEVR
|
| 10 |
+
- synthetic
|
| 11 |
+
- computer-vision
|
| 12 |
+
pretty_name: EditCLEVR Phase 1
|
| 13 |
+
size_categories:
|
| 14 |
+
- 10K<n<100K
|
| 15 |
+
---
|
| 16 |
+
|
| 17 |
+
# EditCLEVR Phase 1
|
| 18 |
+
|
| 19 |
+
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.
|
| 20 |
+
|
| 21 |
+
## Dataset summary
|
| 22 |
+
|
| 23 |
+
- **~20k paired edits** across six evaluation splits
|
| 24 |
+
- **Suites:** atomic single-factor edits, no-edit controls, hard distractors, and CoGenT-OOD combinations
|
| 25 |
+
- **Per-object factors:** `color`, `material`, `size`, `shape`
|
| 26 |
+
- **Artifacts per pair:** before/after RGB images, instance masks (`.npz`), scene JSON, object attributes, edit metadata, and difficulty tags
|
| 27 |
+
|
| 28 |
+
## Splits
|
| 29 |
+
|
| 30 |
+
| Split | Purpose |
|
| 31 |
+
|-------|---------|
|
| 32 |
+
| `train` | Probe training |
|
| 33 |
+
| `val` | Validation |
|
| 34 |
+
| `test_id` | In-distribution atomic edits |
|
| 35 |
+
| `test_noop` | No-edit control pairs |
|
| 36 |
+
| `test_hard` | Hard distractor edits |
|
| 37 |
+
| `test_cogent` | CoGenT-OOD edits |
|
| 38 |
+
|
| 39 |
+
## Download
|
| 40 |
+
|
| 41 |
+
```bash
|
| 42 |
+
pip install huggingface_hub
|
| 43 |
+
git clone https://github.com/torux-bughunter/EditCLEVR-icml.git
|
| 44 |
+
cd EditCLEVR-icml
|
| 45 |
+
bash scripts/setup_data_hf.sh
|
| 46 |
+
```
|
| 47 |
+
|
| 48 |
+
Or from Python:
|
| 49 |
+
|
| 50 |
+
```python
|
| 51 |
+
from editclevr.huggingface import download_dataset
|
| 52 |
+
|
| 53 |
+
download_dataset("outputs/real_phase1_20k_regenerated")
|
| 54 |
+
```
|
| 55 |
+
|
| 56 |
+
## Evaluate
|
| 57 |
+
|
| 58 |
+
```bash
|
| 59 |
+
pip install -r requirements.txt
|
| 60 |
+
python3 -m editclevr.evaluation.run_evaluation \
|
| 61 |
+
--dataset-dir outputs/real_phase1_20k_regenerated \
|
| 62 |
+
--output-dir outputs/evaluation_results
|
| 63 |
+
```
|
| 64 |
+
|
| 65 |
+
## File layout
|
| 66 |
+
|
| 67 |
+
```text
|
| 68 |
+
splits.json
|
| 69 |
+
atomic_id/
|
| 70 |
+
no_edit/
|
| 71 |
+
hard_distractor/
|
| 72 |
+
cogent_ood/
|
| 73 |
+
validation_report.json
|
| 74 |
+
phase1_manifest.json
|
| 75 |
+
```
|
| 76 |
+
|
| 77 |
+
`splits.json` stores relative paths to images, masks, and scene JSON files so the dataset can be moved across machines.
|
| 78 |
+
|
| 79 |
+
## License
|
| 80 |
+
|
| 81 |
+
- **Dataset:** [CC-BY-4.0](https://creativecommons.org/licenses/by/4.0/)
|
| 82 |
+
- **Code:** [MIT](https://github.com/torux-bughunter/EditCLEVR-icml/blob/main/LICENSE)
|
| 83 |
+
|
| 84 |
+
## Citation
|
| 85 |
+
|
| 86 |
+
If you use this dataset, please cite the EditCLEVR ICML workshop paper when available. See [`CITATION.cff`](https://github.com/torux-bughunter/EditCLEVR-icml/blob/main/CITATION.cff) in the code repository.
|
| 87 |
+
|
| 88 |
+
## Source release
|
| 89 |
+
|
| 90 |
+
This mirror corresponds to GitHub release tag `v1.0-data-regenerated-2026-05-19` from the EditCLEVR project.
|