SimVerse / voi
Text-VOI spatial logic puzzle: choose how to rotate and translate base shapes onto a grid so that their XOR-overlapped union exactly reproduces a target pattern.
- Records: 600 levels
- Modality: target pattern image + one image per available base shape
- Output:
{"placements": [{"shape", "angle", "vertex", "grid"}, ...]}
Loading
from datasets import load_dataset
ds = load_dataset("SimVer-ano/simverse2026", "voi")
example = ds["test"][0]
system_text = example["prompt"]["system"]
user_text = example["prompt"]["user"]
target_img = example["images_relative_to_config"]["target"] # e.g. "data/voi-000/target.png"
shape_imgs = example["images_relative_to_config"]["shapes"] # e.g. {"S1": "data/voi-000/shapes/S1.png", ...}
gold_placements = example["answer"]["placements"]
Schema
| Field | Type | Description |
|---|---|---|
ID |
string | Level id, e.g. "voi-000" |
__sample_id__ |
string | Same as ID |
prompt.system / prompt.user |
string | Exact prompt text |
gridSize |
int | Square grid side length |
inventory |
dict | {shape_id: {V1: [x,y], V2: [x,y], ...}} — each base shape's vertex coordinates |
target |
list of polygon dicts | The goal silhouette (XOR of these polygons) |
meta |
dict | difficultyLabel, requiredShapeCount, distractorShapeCount, overlapAllowed |
imageAssets.target |
string | Target pattern image path (level-relative) |
imageAssets.shapes |
dict | {shape_id: image_path} per available shape |
images_relative_to_config |
dict | Same image paths but rewritten to be relative to the config root |
answer.placements |
list[{shape, angle, vertex, grid}] | Reference solution |
legacy_answer |
string | Pre-v1 plain-text DSL form of the answer |
solutionText |
string | The DSL form of the answer (kept because the pixel engine consumes it for the reference mask) |
Output format details
Each placement encodes: rotate shape clockwise by angle ∈ {0,90,180,270} around its local origin, then translate so the post-rotation vertex lands at grid coordinate [gridX, gridY].
The benchmark's pixel engine rasterizes the placements and XOR-combines them; a perfect score requires the resulting mask to equal the target pattern's mask.
License
MIT — see LICENSE at the repo root.