ShawnLi02/JigShape-Train

JigShape: Train & Evaluation Splits

This repository contains the training and evaluation splits of the JigShape benchmark for fine-tuning and evaluating Vision-Language Models on geometric jigsaw puzzle solving.

The held-out test split is available separately at ShawnLi02/JigShape.

Overview

JigShape is a benchmark that evaluates joint visual-geometric reasoning in VLMs. Unlike traditional jigsaw benchmarks that use rectangular cuts (which create ambiguous ground truth in repeated-texture regions), JigShape features tab-and-blank interlocking pieces where geometric constraints ensure every puzzle has a unique solution.

Models must predict the correct grid position for each labeled piece by reasoning about both visual content (texture, color, object boundaries) and geometric constraints (tab-blank edge compatibility).

Dataset Statistics

Split	4x4	8x8	12x12	16x16	Total
Train	22,992	22,992	22,992	22,992	91,968
Eval	250	250	250	250	1,000

• Source images: 23,742 unique high-resolution images from DIV2K, DIV8K, and Unsplash
• No overlap: Train and eval splits are partitioned by source image; the same image never appears in both

Directory Structure

train/
  grid_4x4/
    DIV2K_0001/
      layout.png          # Shuffled pieces displayed on a grid, labeled with piece IDs
      source.png          # Original image (ground truth reference)
      ground_truth.json   # Piece-to-position mapping and edge signatures
    DIV2K_0002/
    ...
  grid_8x8/
  grid_12x12/
  grid_16x16/

validation/
  grid_4x4/
  grid_8x8/
  grid_12x12/
  grid_16x16/

split_index.json          # Canonical list of image IDs per split

Instance Format

Each instance directory contains three files:

layout.png

The model input: all N x N pieces arranged in ID order (not solution order) on a display board. Each piece is labeled with its numeric ID and shows its tab/blank/flat edge shapes.

source.png

The original uncut image, provided for reference and visualization.

ground_truth.json

{
  "instance_id": "DIV2K_0001",
  "grid_size": 4,
  "n_pieces": 16,
  "id_to_position": {
    "7": [0, 0],
    "9": [0, 1],
    "5": [0, 2],
    "...": "..."
  },
  "edge_signatures": {
    "7": {"top": "flat", "right": "tab", "bottom": "blank", "left": "flat"},
    "9": {"top": "flat", "right": "tab", "bottom": "tab", "left": "blank"},
    "...": "..."
  }
}

Fields:

• id_to_position: Maps each piece ID to its correct [row, col] position in the solved puzzle
• edge_signatures: Each piece's four edges typed as tab (convex), blank (concave), or flat (border)

Edge Types & Compatibility Rules

Edge Type	Description	Constraint
Tab	Convex semicircular protrusion	Must pair with a blank on the adjacent piece
Blank	Concave semicircular indentation	Must pair with a tab on the adjacent piece
Flat	Straight edge	Only on puzzle borders (corners have 2, edges have 1, interior pieces have 0)

Evaluation Metrics

Metric	Description
Piece Accuracy (PA)	Fraction of pieces placed in their correct position
Exact Match (EM)	Whether the entire puzzle is solved correctly (all pieces correct)

Quick Start

from huggingface_hub import snapshot_download

# Download eval split only (~5 GB)
snapshot_download(
    repo_id="ShawnLi02/JigShape-Train",
    repo_type="dataset",
    allow_patterns="validation/**",
    local_dir="./JigShape"
)

# Download a specific grid size for training (~120 GB for 16x16)
snapshot_download(
    repo_id="ShawnLi02/JigShape-Train",
    repo_type="dataset",
    allow_patterns="train/grid_4x4/**",
    local_dir="./JigShape"
)

Related Resources

• Test split (held-out for competition): ShawnLi02/JigShape
• Paper: JigShape: Can Vision-Language Models Solve Jigsaw Puzzles? (under review)

Citation

@article{jigshape2025,
  title={JigShape: Can Vision-Language Models Solve Jigsaw Puzzles?},
  author={Anonymous},
  year={2025}
}

License

This dataset is released under the CC BY 4.0 license.