SimVerse / cube2
Cube goal-roll (top-face directions): given a cube whose initial outer-surface configuration is shown as a cross net and a target top-face image, output a sequence of N/S/E/W rolls so the cube's top face matches the target.
- Records: 502 levels
- Modality: initial cross-net image + target top-face image
- Output:
{"directions": ["N", "E", "S", ...]}(open-ended; multiple sequences may be valid)
Loading
from datasets import load_dataset
ds = load_dataset("SimVer-ano/simverse2026", "cube2")
example = ds["test"][0]
system_text = example["prompt"]["system"]
user_text = example["prompt"]["user"]
initial_img = example["images_relative_to_config"]["initialNetImage"] # e.g. "images/C001/initial_net.png"
target_img = example["images_relative_to_config"]["targetTopFaceImage"] # e.g. "images/C001/target_top_face.png"
reference_directions = example["answer"]["directions"] # one known-valid sequence, NOT the only valid one
Schema
| Field | Type | Description |
|---|---|---|
code |
string | Level id, e.g. "C001" |
__sample_id__ |
string | Same as code |
prompt.system / prompt.user |
string | Exact prompt text |
taskType |
string | Always "roll_to_target_top_face" |
initialCube.net.cells |
list | The visible cross-net cells with (faceKey, row, col, patternId, rotation) |
initialCube.solutionFaces |
dict | Visible faces' (patternId, rotation) mapping |
targetTopFace |
dict | Target's patternId, rotation |
rollSequence |
list[string] | The original roll sequence used to generate this puzzle |
observedPathFaces |
list | Per-step bottom-face imprints (mostly informational) |
metadata |
dict | difficulty, tier, moveCount, etc. |
imagePaths |
dict | initialNetImage, targetTopFaceImage (relative to data dir) |
images_relative_to_config |
dict | Same paths rewritten to be relative to this config's root |
answer.directions |
list[string] | One known-valid direction sequence; the validator accepts any sequence whose simulated top-face matches the target |
legacy_answer |
dict | Pre-v1 face-map answer from the original cube reconstruction task variant (kept for back-compat) |
Open-ended scoring
cube2 is open-ended: the reference answer.directions is just one known-valid sequence. The benchmark's validator runs the engine on the model's directions and checks whether the resulting top face matches targetTopFace. Multiple distinct sequences can earn full credit.
Companion file
catalog.json mirrors data2/index.json — a fat JSON the frontend demo uses for the level grid.
License
MIT — see LICENSE at the repo root.