CurveBench-Synthetic

CurveBench-Synthetic is a procedurally generated companion to CurveBench.
Each sample is an image of non-crossing closed curves paired with the rooted containment tree that encodes nesting (root 0 = exterior).

This set isolates two synthetic geometry regimes:

Split	Size	Description
convex	50	Polygon-style / more convex closed boundaries
non_convex	50	Organic / composed non-convex closed boundaries
combined	100	Both regimes merged

---

Fields

Field	Type	Description
image	Image	Synthetic curve image (PNG)
category	string	"Convex" or "Non-convex"
filename	string	Original filename
num_nodes	int32	Number of non-root regions / tree edges
tree	string	Stringified edge list, e.g. "[(0, 1), (1, 2)]" — each tuple is (parent, child)
style	string	Generator style tag (e.g. polygon, compose-smooth)

Example

from datasets import load_dataset

ds = load_dataset("AmirMohseni/CurveBench-Synthetic", split="combined")
row = ds[0]
print(row["num_nodes"], row["tree"], row["category"])
row["image"]  # PIL.Image

---

Task format

Same answer format as CurveBench:

<answer>
N
u v
...
</answer>

• N = number of non-root nodes
• each line u v means edge from parent v to child u
• root is always 0 (exterior)

Scoring uses rooted tree isomorphism (unordered siblings; root fixed at 0).

---

Relation to CurveBench

• AmirMohseni/CurveBench-Easy — hand-drawn Easy (≤5 curves)
• AmirMohseni/CurveBench — hand-drawn Hard
• This dataset — synthetic Convex / Non-convex for controlled geometry ablations

---

Citation

If you use CurveBench or CurveBench-Synthetic, please cite the CurveBench paper / collection.