Update README.md
Browse files
README.md
CHANGED
|
@@ -1,3 +1,248 @@
|
|
| 1 |
---
|
| 2 |
license: cc-by-4.0
|
| 3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
---
|
| 2 |
license: cc-by-4.0
|
| 3 |
---
|
| 4 |
+
|
| 5 |
+
# CREBench Reward Data
|
| 6 |
+
|
| 7 |
+
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.
|
| 8 |
+
|
| 9 |
+
**Hugging Face:** [hanjy/CREBench](https://huggingface.co/datasets/hanjy/CREBench)
|
| 10 |
+
|
| 11 |
+
**Local root (after download):** `./CREBench/`
|
| 12 |
+
|
| 13 |
+
| Benchmark | Path | Label source |
|
| 14 |
+
|-----------|------|----------------|
|
| 15 |
+
| **CREBench_LVLM** | `CREBench_LVLM/` | Gemma-3-27B-IT pairwise judgments |
|
| 16 |
+
| **CREBench_Human** | `CREBench_Human/` | 5 human annotators per pair |
|
| 17 |
+
|
| 18 |
+
---
|
| 19 |
+
|
| 20 |
+
## Object categories
|
| 21 |
+
|
| 22 |
+
Both benchmarks use the same five object categories:
|
| 23 |
+
|
| 24 |
+
- `chair`
|
| 25 |
+
- `bowl`
|
| 26 |
+
- `car`
|
| 27 |
+
- `vase`
|
| 28 |
+
- `handbag`
|
| 29 |
+
|
| 30 |
+
---
|
| 31 |
+
|
| 32 |
+
## Directory layout
|
| 33 |
+
|
| 34 |
+
```
|
| 35 |
+
CREBench/
|
| 36 |
+
├── CREBench_LVLM/
|
| 37 |
+
│ ├── Label/
|
| 38 |
+
│ │ ├── gemma27b_preferences_chair.pkl
|
| 39 |
+
│ │ ├── gemma27b_preferences_bowl.pkl
|
| 40 |
+
│ │ ├── gemma27b_preferences_car.pkl
|
| 41 |
+
│ │ ├── gemma27b_preferences_vase.pkl
|
| 42 |
+
│ │ └── gemma27b_preferences_handbag.pkl
|
| 43 |
+
│ ├── chair/ # pair_0000_0.png, pair_0000_1.png, ...
|
| 44 |
+
│ ├── bowl/
|
| 45 |
+
│ ├── car/
|
| 46 |
+
│ ├── vase/
|
| 47 |
+
│ └── handbag/
|
| 48 |
+
└── CREBench_Human/
|
| 49 |
+
├── Label/
|
| 50 |
+
│ ├── human_preferences_chair.pkl
|
| 51 |
+
│ ├── human_preferences_bowl.pkl
|
| 52 |
+
│ ├── human_preferences_car.pkl
|
| 53 |
+
│ ├── human_preferences_vase.pkl
|
| 54 |
+
│ └── human_preferences_handbag.pkl
|
| 55 |
+
├── chair/
|
| 56 |
+
├── bowl/
|
| 57 |
+
├── car/
|
| 58 |
+
├── vase/
|
| 59 |
+
└── handbag/
|
| 60 |
+
```
|
| 61 |
+
|
| 62 |
+
---
|
| 63 |
+
|
| 64 |
+
## Image pairs
|
| 65 |
+
|
| 66 |
+
Each sample is a pair of images shown side-by-side during annotation.
|
| 67 |
+
|
| 68 |
+
| Field | Description |
|
| 69 |
+
|-------|-------------|
|
| 70 |
+
| **Naming** | `pair_{idx:04d}_0.png` and `pair_{idx:04d}_1.png` |
|
| 71 |
+
| **Index** | `idx` runs from `0` to `N-1` within each category folder |
|
| 72 |
+
| **Image `_0`** | First image in the pair (option A) |
|
| 73 |
+
| **Image `_1`** | Second image in the pair (option B) |
|
| 74 |
+
|
| 75 |
+
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`.
|
| 76 |
+
|
| 77 |
+
### Dataset sizes
|
| 78 |
+
|
| 79 |
+
| Benchmark | Pairs per category | Total pairs | Total images |
|
| 80 |
+
|-----------|-------------------|-------------|--------------|
|
| 81 |
+
| **CREBench_LVLM** | 1,000 | 5,000 | 10,000 |
|
| 82 |
+
| **CREBench_Human** | 100 | 500 | 1,000 |
|
| 83 |
+
|
| 84 |
+
---
|
| 85 |
+
|
| 86 |
+
## Labels (`.pkl`)
|
| 87 |
+
|
| 88 |
+
Labels are stored as NumPy arrays inside pickle files (`numpy.ndarray`, `dtype=int64`).
|
| 89 |
+
|
| 90 |
+
### Label dimensions (columns)
|
| 91 |
+
|
| 92 |
+
Each row has **4** integer labels in this order:
|
| 93 |
+
|
| 94 |
+
| Column | Dimension |
|
| 95 |
+
|--------|-----------|
|
| 96 |
+
| 0 | **Geometry** creativity |
|
| 97 |
+
| 1 | **Material** creativity |
|
| 98 |
+
| 2 | **Texture** creativity |
|
| 99 |
+
| 3 | **Overall** creativity |
|
| 100 |
+
|
| 101 |
+
### Label values
|
| 102 |
+
|
| 103 |
+
All entries are in **`{-1, 0, 1}`**:
|
| 104 |
+
|
| 105 |
+
| Value | Meaning |
|
| 106 |
+
|-------|---------|
|
| 107 |
+
| **`0`** | `pair_*_0.png` is more creative on that dimension |
|
| 108 |
+
| **`1`** | `pair_*_1.png` is more creative on that dimension |
|
| 109 |
+
| **`-1`** | Tie, no clear preference, or not decidable |
|
| 110 |
+
|
| 111 |
+
### Shapes
|
| 112 |
+
|
| 113 |
+
| Benchmark | File pattern | Shape | Description |
|
| 114 |
+
|-----------|--------------|-------|-------------|
|
| 115 |
+
| **CREBench_LVLM** | `Label/gemma27b_preferences_{category}.pkl` | **`(N, 4)`** | `N` = number of pairs in that category (1,000) |
|
| 116 |
+
| **CREBench_Human** | `Label/human_preferences_{category}.pkl` | **`(5, N, 4)`** | 5 annotators × `N` pairs (100) × 4 dimensions |
|
| 117 |
+
|
| 118 |
+
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.
|
| 119 |
+
|
| 120 |
+
---
|
| 121 |
+
|
| 122 |
+
## CREBench_LVLM
|
| 123 |
+
|
| 124 |
+
LVLM preferences from **Gemma-3-27B-IT** (`google/gemma-3-27b-it`) on pairwise image comparisons.
|
| 125 |
+
|
| 126 |
+
### Annotation protocol
|
| 127 |
+
|
| 128 |
+
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`).
|
| 129 |
+
|
| 130 |
+
### Example (chair)
|
| 131 |
+
|
| 132 |
+
- **Images:** `CREBench_LVLM/chair/pair_0000_0.png`, `pair_0000_1.png`, …, `pair_0999_1.png`
|
| 133 |
+
- **Labels:** `CREBench_LVLM/Label/gemma27b_preferences_chair.pkl` with shape `(1000, 4)`
|
| 134 |
+
|
| 135 |
+
---
|
| 136 |
+
|
| 137 |
+
## CREBench_Human
|
| 138 |
+
|
| 139 |
+
Human pairwise creativity judgments on a smaller subset of pairs (100 per category).
|
| 140 |
+
|
| 141 |
+
### Annotation protocol
|
| 142 |
+
|
| 143 |
+
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`).
|
| 144 |
+
|
| 145 |
+
### Example (chair)
|
| 146 |
+
|
| 147 |
+
- **Images:** `CREBench_Human/chair/pair_0000_0.png`, …, `pair_0099_1.png`
|
| 148 |
+
- **Labels:** `CREBench_Human/Label/human_preferences_chair.pkl` with shape `(5, 100, 4)`
|
| 149 |
+
|
| 150 |
+
---
|
| 151 |
+
|
| 152 |
+
## Download and usage in Python
|
| 153 |
+
|
| 154 |
+
Install [huggingface_hub](https://huggingface.co/docs/huggingface_hub) if needed:
|
| 155 |
+
|
| 156 |
+
```bash
|
| 157 |
+
pip install huggingface_hub
|
| 158 |
+
```
|
| 159 |
+
|
| 160 |
+
### Download the full dataset
|
| 161 |
+
|
| 162 |
+
The Hub repo [`hanjy/CREBench`](https://huggingface.co/datasets/hanjy/CREBench) contains `CREBench_LVLM/` and `CREBench_Human/` at the top level:
|
| 163 |
+
|
| 164 |
+
```python
|
| 165 |
+
from huggingface_hub import snapshot_download
|
| 166 |
+
|
| 167 |
+
ROOT = snapshot_download(
|
| 168 |
+
repo_id="hanjy/CREBench",
|
| 169 |
+
repo_type="dataset",
|
| 170 |
+
local_dir="./CREBench",
|
| 171 |
+
)
|
| 172 |
+
# ./CREBench/CREBench_LVLM/
|
| 173 |
+
# ./CREBench/CREBench_Human/
|
| 174 |
+
```
|
| 175 |
+
|
| 176 |
+
Equivalent CLI:
|
| 177 |
+
|
| 178 |
+
```bash
|
| 179 |
+
huggingface-cli download hanjy/CREBench \
|
| 180 |
+
--repo-type dataset \
|
| 181 |
+
--local-dir ./CREBench
|
| 182 |
+
```
|
| 183 |
+
|
| 184 |
+
For a private or gated repo, log in first (`huggingface-cli login` or set the `HF_TOKEN` environment variable).
|
| 185 |
+
|
| 186 |
+
### Download a single file
|
| 187 |
+
|
| 188 |
+
```python
|
| 189 |
+
from huggingface_hub import hf_hub_download
|
| 190 |
+
|
| 191 |
+
label_path = hf_hub_download(
|
| 192 |
+
repo_id="hanjy/CREBench",
|
| 193 |
+
repo_type="dataset",
|
| 194 |
+
filename="CREBench_LVLM/Label/gemma27b_preferences_chair.pkl",
|
| 195 |
+
local_dir="./CREBench",
|
| 196 |
+
)
|
| 197 |
+
```
|
| 198 |
+
|
| 199 |
+
### Load labels and images
|
| 200 |
+
|
| 201 |
+
```python
|
| 202 |
+
import pickle
|
| 203 |
+
from pathlib import Path
|
| 204 |
+
|
| 205 |
+
ROOT = Path("./CREBench")
|
| 206 |
+
category = "chair"
|
| 207 |
+
|
| 208 |
+
# CREBench_LVLM — (num_samples, 4)
|
| 209 |
+
with open(ROOT / "CREBench_LVLM/Label" / f"gemma27b_preferences_{category}.pkl", "rb") as f:
|
| 210 |
+
lvlm_labels = pickle.load(f) # e.g. shape (1000, 4)
|
| 211 |
+
|
| 212 |
+
# CREBench_Human — (5, num_samples, 4)
|
| 213 |
+
with open(ROOT / "CREBench_Human/Label" / f"human_preferences_{category}.pkl", "rb") as f:
|
| 214 |
+
human_labels = pickle.load(f) # e.g. shape (5, 100, 4)
|
| 215 |
+
|
| 216 |
+
geom, mat, tex, overall = 0, 1, 2, 3
|
| 217 |
+
pair_idx = 0
|
| 218 |
+
print("LVLM overall preference:", lvlm_labels[pair_idx, overall])
|
| 219 |
+
print("Human overall (5 raters):", human_labels[:, pair_idx, overall])
|
| 220 |
+
```
|
| 221 |
+
|
| 222 |
+
### Pairing images with labels
|
| 223 |
+
|
| 224 |
+
```python
|
| 225 |
+
def pair_paths(root: Path, category: str, idx: int):
|
| 226 |
+
d = root / category
|
| 227 |
+
return d / f"pair_{idx:04d}_0.png", d / f"pair_{idx:04d}_1.png"
|
| 228 |
+
|
| 229 |
+
img0, img1 = pair_paths(ROOT / "CREBench_LVLM", "chair", 0)
|
| 230 |
+
label_row = lvlm_labels[0] # [geometry, material, texture, overall]
|
| 231 |
+
```
|
| 232 |
+
|
| 233 |
+
|
| 234 |
+
---
|
| 235 |
+
|
| 236 |
+
## Citation
|
| 237 |
+
|
| 238 |
+
If you use CREBench_LVLM or CREBench_Human, please cite the CREward paper:
|
| 239 |
+
|
| 240 |
+
```bibtex
|
| 241 |
+
@inproceedings{han2026creward,
|
| 242 |
+
title={CREward: A Type-Specific Creativity Reward Model},
|
| 243 |
+
author={Han, Jiyeon and Mahdavi-Amiri, Ali and Zhang, Hao and Jeong, Haedong},
|
| 244 |
+
booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
|
| 245 |
+
pages={21932--21941},
|
| 246 |
+
year={2026}
|
| 247 |
+
}
|
| 248 |
+
```
|