Dataset/combination_dataset_3000/README.md

2D Combination Dataset (3,000 images)

1. Dataset purpose

This synthetic dataset implements the 2D Combination task in the Spatial-DISE taxonomy. It tests part-whole spatial composition: mentally rotating and placing smaller pieces to reconstruct a target. This differs from transformation matching, which compares whole shapes, and cube-net folding, which requires 2D-to-3D reasoning. The taxonomy reference is Huang et al., Spatial-DISE: A Unified Benchmark for Evaluating Spatial Reasoning in Vision-Language Models, arXiv:2510.13394.

2. Visual design

Each 550–600 × 400–450 PNG uses a dark #1A1A1A background and one light outline color. A connected 6–10-cell target appears above four labeled candidate sets. Visible unit-cell boundaries make area and piece shape inspectable without color hints. Each candidate contains 2–4 connected polyomino pieces.

3. Candidate construction and reasoning skills

Every image has exactly one rotation-and-translation solution and exactly three verified distractors:

• equal area but no exact cover, even when reflection is allowed;
• wrong total area;
• equal area and solvable with reflection, but not with rotation alone.

The tasks exercise part-whole composition, mental piece rotation and placement, area-based elimination, and rejection of plausible mirrored or non-fitting assemblies.

4. Questions and annotations

annotations.jsonl contains one record per image with target_cells, target_cell_count, four candidate piece sets, failure reasons, the correct choice, seed, canvas size, and four ordered questions:

1. target cell count;
2. the core exact-assembly choice;
3. candidate piece count or candidate/target area comparison;
4. reflection distractor, wrong-area distractor, or attached-cell connectivity.

The flattened dataset_final.csv and dataset_final.jsonl contain exactly 12,000 question rows using columns task, image, prompt, groundtruth, and compact metadata.

5. Generation and reproducibility

Python 3 and Pillow are required. Image i uses random.Random(i), so generation is reproducible.

python generate_combination_dataset.py --n 3000 --output-dir .
python flatten_annotations.py --dataset-dir .
python make_contact_sheet.py --dataset-dir . --output contact_sheet.png --count 20
python make_contact_sheet.py --dataset-dir . --output human_calibration/review_sheet.png --count 5 --columns 5

For a five-image development run, use python generate_combination_dataset.py --sample --output-dir sample_output.

6. Validation

Run:

python validate_combination_dataset.py --dataset-dir . --report validation_report.txt

The validator is independent of the generator. It reconstructs rotations, reflections, translations, and exact-cover search; checks each declared failure reason; recomputes all question answers; verifies A/B/C/D answer balance; confirms files and sizes; and reads every PNG to check that expected target and piece cell boundaries were actually rendered. A successful full run reports zero mismatches.

7. Known limitations

Shapes are restricted to unit-grid polyominoes for exact verification rather than free-form polygons. Candidate sets contain only 2–4 pieces. Reflection is strictly invalid for the core question, with no partial credit for an almost-correct mirrored construction. The images test spatial composition but do not model physical thickness, material, or real-world assembly constraints.

Supplementary open-ended questions

Version combination-4.0.0 replaces the supplementary free-response set with the exact domain template in OPEN_QUESTION_SPEC.md. It includes 3000 eligible items and excludes 0 under {}. Every prompt uses visible evidence, asks for justification, and ends with a confidence score from 0 to 1.

• open_questions.csv is public and contains exactly question_id,image,prompt.
• open_answer_key.csv is answer-key-side and contains separate partial-credit fields, an exhaustive acceptance_set, deterministic targets, and machine-readable tolerances for every numeric field.
• open_annotations.jsonl is answer-key-side and adds the complete derivation and scoring declaration.
• open_validation_metrics.json contains full target and answer distributions, constant-answer baselines, prompt/schema checks, independent metadata derivation results, and exhaustive PNG recovery results.

Scored fields at or above 60% after template revision: {}. Targeted fields: correct_candidate, rejected_candidate, rejection_reason. Do not provide open_answer_key.csv, open_annotations.jsonl, or the closed-set annotations.jsonl to a model under evaluation because they expose answer-side scene metadata.