CloudAVR Training Puzzles

This dataset contains the point-cloud puzzles used for the CloudAVR low-data training run. It is provided as sample puzzles and dataset documentation; this repository intentionally does not include the project source code or model checkpoints.

Dataset summary

• 7,500 procedurally generated 3×3 matrix-completion puzzles.
• 6 rule templates: Conservation, Count, Distribute-three, Permutation, Progression, and Symmetry.
• 1,250 puzzles per rule: 1,000 training examples and 250 validation examples under the experiment's stratified 80/20 split. There is no internal test split in this subset.
• Each puzzle has one missing matrix position and four candidate point clouds, with one correct candidate.
• Point clouds are ASCII PLY files containing XYZ coordinates and RGB colors.

Repository layout

meta.json
<RuleName>/meta.json
<RuleName>/sample_XXXXXX/meta.json
<RuleName>/sample_XXXXXX/grid_<row>_<col>.ply
<RuleName>/sample_XXXXXX/cand_<index>.ply

The root meta.json contains the complete annotation list. Each rule directory also contains its rule-specific metadata. Paths in each metadata record are relative to the dataset root.

Annotation fields

A sample metadata record includes the observed-grid mask, grid and candidate PLY paths, the ground-truth candidate index/label, the rule template and parameters, matrix relations, entity attributes, and distractor descriptions. The generated metadata exposes the intended transformation so the set can be used for controlled analysis of visual and geometric reasoning.

Implementation details

The task is formulated as point-cloud matrix reasoning. A puzzle consists of a 3×3 grid of structured point clouds with one cell hidden. The remaining cells define a transformation pattern governed by one rule instance; the model must select the correct completion from four candidates. The training implementation reads the PLY XYZ/RGB attributes, normalizes and resamples point clouds to a common input size, computes global geometric features, and performs a stratified split by rule template with a fixed seed. The details needed to reproduce the data interpretation are included in the metadata; source code is intentionally not released here.

Intended use

For research on 3D/point-cloud matrix reasoning, controlled rule generalization, candidate discrimination, and robustness to geometric distractors. Please inspect the metadata and the applicable project terms before redistribution or commercial use.