CREBench Reward Data

Pairwise creativity preference benchmarks used for CREward training and evaluation. Each benchmark contains rendered image pairs and human- or LVLM-derived preference labels along four creativity dimensions.

Hugging Face: hanjy/CREBench

Local root (after download): ./CREBench/

Benchmark	Path	Label source
CREBench_LVLM	CREBench_LVLM/	Gemma-3-27B-IT pairwise judgments
CREBench_Human	CREBench_Human/	5 human annotators per pair

---

Object categories

Both benchmarks use the same five object categories:

• chair
• bowl
• car
• vase
• handbag

---

Directory layout

CREBench/
├── CREBench_LVLM/
│   ├── Label/
│   │   ├── gemma27b_preferences_chair.pkl
│   │   ├── gemma27b_preferences_bowl.pkl
│   │   ├── gemma27b_preferences_car.pkl
│   │   ├── gemma27b_preferences_vase.pkl
│   │   └── gemma27b_preferences_handbag.pkl
│   ├── chair/          # pair_0000_0.png, pair_0000_1.png, ...
│   ├── bowl/
│   ├── car/
│   ├── vase/
│   └── handbag/
└── CREBench_Human/
    ├── Label/
    │   ├── human_preferences_chair.pkl
    │   ├── human_preferences_bowl.pkl
    │   ├── human_preferences_car.pkl
    │   ├── human_preferences_vase.pkl
    │   └── human_preferences_handbag.pkl
    ├── chair/
    ├── bowl/
    ├── car/
    ├── vase/
    └── handbag/

---

Image pairs

Each sample is a pair of images shown side-by-side during annotation.

Field	Description
Naming	pair_{idx:04d}_0.png and pair_{idx:04d}_1.png
Index	idx runs from 0 to N-1 within each category folder
Image _0	First image in the pair (option A)
Image _1	Second image in the pair (option B)

Pairs are aligned with label rows by sorted filename order: row i corresponds to pair_{i:04d}_0.png and pair_{i:04d}_1.png.

Dataset sizes

Benchmark	Pairs per category	Total pairs	Total images
CREBench_LVLM	1,000	5,000	10,000
CREBench_Human	100	500	1,000

---

Labels (.pkl)

Labels are stored as NumPy arrays inside pickle files (numpy.ndarray, dtype=int64).

Label dimensions (columns)

Each row has 4 integer labels in this order:

Column	Dimension
0	Geometry creativity
1	Material creativity
2	Texture creativity
3	Overall creativity

Label values

All entries are in {-1, 0, 1}:

Value	Meaning
0	pair_*_0.png is more creative on that dimension
1	pair_*_1.png is more creative on that dimension
-1	Tie, no clear preference, or not decidable

Shapes

Benchmark	File pattern	Shape	Description
CREBench_LVLM	Label/gemma27b_preferences_{category}.pkl	(N, 4)	N = number of pairs in that category (1,000)
CREBench_Human	Label/human_preferences_{category}.pkl	(5, N, 4)	5 annotators × N pairs (100) × 4 dimensions

So for CREBench_LVLM, labels are num_samples × 4. For CREBench_Human, raw per-rater labels are 5 × num_samples × 4; aggregate across raters (e.g. majority vote or mean) if a single preference vector per pair is needed.

---

CREBench_LVLM

LVLM preferences from Gemma-3-27B-IT (google/gemma-3-27b-it) on pairwise image comparisons.

Annotation protocol

For each pair, the model answers four questions (geometry, material, texture, overall), choosing A, B, or Not decidable. Responses are converted to the signed label format above (A → 0 for image _0, B → 1 for image _1, not decidable → -1).

Example (chair)

• Images: CREBench_LVLM/chair/pair_0000_0.png, pair_0000_1.png, …, pair_0999_1.png
• Labels: CREBench_LVLM/Label/gemma27b_preferences_chair.pkl with shape (1000, 4)

---

CREBench_Human

Human pairwise creativity judgments on a smaller subset of pairs (100 per category).

Annotation protocol

Five raters judge each pair on the same four dimensions (geometry, material, texture, overall). Labels are stored per rater in the first dimension of the array (shape[0] == 5).

Example (chair)

• Images: CREBench_Human/chair/pair_0000_0.png, …, pair_0099_1.png
• Labels: CREBench_Human/Label/human_preferences_chair.pkl with shape (5, 100, 4)

---

Download and usage in Python

Install huggingface_hub if needed:

pip install huggingface_hub

Download the full dataset

The Hub repo hanjy/CREBench contains CREBench_LVLM/ and CREBench_Human/ at the top level:

from huggingface_hub import snapshot_download

ROOT = snapshot_download(
    repo_id="hanjy/CREBench",
    repo_type="dataset",
    local_dir="./CREBench",
)
# ./CREBench/CREBench_LVLM/
# ./CREBench/CREBench_Human/

Equivalent CLI:

huggingface-cli download hanjy/CREBench \
  --repo-type dataset \
  --local-dir ./CREBench

For a private or gated repo, log in first (huggingface-cli login or set the HF_TOKEN environment variable).

Download a single file

from huggingface_hub import hf_hub_download

label_path = hf_hub_download(
    repo_id="hanjy/CREBench",
    repo_type="dataset",
    filename="CREBench_LVLM/Label/gemma27b_preferences_chair.pkl",
    local_dir="./CREBench",
)

Load labels and images

import pickle
from pathlib import Path

ROOT = Path("./CREBench")
category = "chair"

# CREBench_LVLM — (num_samples, 4)
with open(ROOT / "CREBench_LVLM/Label" / f"gemma27b_preferences_{category}.pkl", "rb") as f:
    lvlm_labels = pickle.load(f)  # e.g. shape (1000, 4)

# CREBench_Human — (5, num_samples, 4)
with open(ROOT / "CREBench_Human/Label" / f"human_preferences_{category}.pkl", "rb") as f:
    human_labels = pickle.load(f)  # e.g. shape (5, 100, 4)

geom, mat, tex, overall = 0, 1, 2, 3
pair_idx = 0
print("LVLM overall preference:", lvlm_labels[pair_idx, overall])
print("Human overall (5 raters):", human_labels[:, pair_idx, overall])

Pairing images with labels

def pair_paths(root: Path, category: str, idx: int):
    d = root / category
    return d / f"pair_{idx:04d}_0.png", d / f"pair_{idx:04d}_1.png"

img0, img1 = pair_paths(ROOT / "CREBench_LVLM", "chair", 0)
label_row = lvlm_labels[0]  # [geometry, material, texture, overall]

---

Citation

If you use CREBench_LVLM or CREBench_Human, please cite the CREward paper:

@inproceedings{han2026creward,
  title={CREward: A Type-Specific Creativity Reward Model},
  author={Han, Jiyeon and Mahdavi-Amiri, Ali and Zhang, Hao and Jeong, Haedong},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={21932--21941},
  year={2026}
}