Dataset/combination3d_dataset_3000/README.md

GRIP 3D Combination Dataset 3000 — v2

Version combination3d-2.0.0 contains 3,000 dark-theme voxel-assembly images and 15,000 questions. Each target has 8–14 cubes and four candidate piece sets: one valid z-spin assembly, one full-rotation-resistant gap/overlap near miss, one wrong-count set, and one set that succeeds only after a 3D tumble.

Explicit coordinate convention

The dataset uses a right-handed (x,y,z) voxel frame. The isometric renderer maps:

screen_x = 0.8660254 * (x - y)
screen_y = 0.5 * (x + y) - z

Therefore z is the visual vertical axis: increasing z moves a cube upward on screen. max_height_layers is the inclusive z extent, and the permitted “rotation around the vertical axis” in Levels 2 and 5 is a z-axis rotation. Every record stores vertical_axis: "z", coordinate_frame, renderer_projection, and permitted_piece_rotation_axes: ["z"] so the convention cannot drift.

Manual-review item combination3d_0017 has y extent 3 and z extent 4; its stored height 4 is correct because the renderer and questions both use z. Across the previous build, zero Level 4 count-and-height answers were wrong.

Five questions

1. Count target cubes.
2. Select the candidate that tiles the target using translation and z-axis spins only.
3. Compare piece count or cube total.
4. Identify a failure class or combine total count with z-height.
5. Apply a whole-structure z rotation or add a cube above the highest z layer.

Exhaustive validation

The validator independently enumerates four z-axis orientations and all 24 proper cube orientations, searches translations and non-overlapping exact covers, and confirms for every item that exactly one candidate succeeds under the permitted transform set. It also independently verifies all four failure reasons. Final results: zero correct-candidate failures, zero successful distractors, and zero failure-reason mismatches.

Every final PNG is checked for target presence, exact candidate piece-panel recovery, canvas dimensions, and the structure signature needed by cube count, piece count, and z-height questions. The report also includes all five answer distributions and constant baselines, the unfiltered stored-feature leak audit, full parameter distributions, reference-frame audit, and guard injection tests. The final v2 build passes 3,000 images and 15,000 questions with zero mismatches.

The superseded metadata and reports are under archive/v1/; the unchanged image renderer is shared by the builds.

Commands

python generate_combination3d_dataset.py --n 3000 --output-dir .
python flatten_annotations.py --dataset-dir .
python validate_combination3d_dataset.py --dataset-dir .

Public evaluation and feature classification

question_set.csv is the only model-facing file and contains exactly question_id, task, image, and prompt. The remaining tabular files are private answer-key-side artifacts; do not expose annotations.jsonl to a tested model.

target_cube_count and the z-derived target height are definitional for Level 4, which explicitly asks for target count and height. They are non-definitional for Level 5 rotation-invariance answers. Their bias-corrected Cramér's V values against Level 5 are 0.8084 and 0.8104 respectively, but both values live only in private answer-key-side annotations and never appear in question_set.csv. The complete post-cleanup matrix is recorded in the suite-level release_hygiene_audit.json.

Supplementary open-ended questions

Version combination3d-5.0.0 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.