metadata
license: cc-by-4.0
pretty_name: COSPLAN
configs:
- config_name: Maze-E
data_files:
- split: test
path: checkboard/metadata.jsonl
- config_name: Blocksworld-E
data_files:
- split: test
path: blocksworld/metadata.jsonl
- config_name: Shuffle-E
data_files:
- split: test
path: imagenet/metadata.jsonl
- config_name: Robovqa-E
data_files:
- split: test
path: robovqa/metadata.jsonl
COSPLAN
A composite visual-planning benchmark with four multiple-choice subsets. Each
subset is a separate config in HuggingFace imagefolder format (images +
metadata.jsonl). Every row has an image, a prompt, and a correct_option
(A/B/C/D).
All four subsets are step-completion tasks: given a partial sequence (with one erroneous step) and an image, select the option that correctly completes the task.
| Config | Items | Task |
|---|---|---|
Maze-E |
5000 | Complete the path from start to goal around obstacles |
Blocksworld-E |
5000 | Complete the block-move sequence to the target configuration |
Shuffle-E |
5000 | Complete the patch-swap sequence to rearrange into the target image |
Robovqa-E |
390 | Complete the robot action sequence |
Usage
from datasets import load_dataset
ds = load_dataset("shrg7/COSPLAN", "Blocksworld-E", split="test")
row = ds[0]
print(row["prompt"], row["correct_option"])
# The image is decoded as a PIL.Image under the "image" column
img = row["image"]
print(img.size, img.mode)
img.save("example.png") # or img.show()