---
pretty_name: LogiPuz Pools
language:
- en
tags:
- synthetic
- logic-puzzles
- reasoning
- benchmark
- zebra-logic
- z3
task_categories:
- question-answering
size_categories:
- 100K<n<1M
license: cc-by-4.0
configs:
- config_name: shape_summary
  default: true
  data_files:
  - split: train
    path: viewer/shape_summary.jsonl
---

# LogiPuz Pools

Reusable core candidate pools for LogiPuz ZebraLogic-style puzzle generation.

This artifact contains theme-free, solver-checked core puzzle candidates for every shape from `2x2` through `6x6`. The pools are intended for dataset construction, support checks, and reproducible extensions of the LogiPuz benchmark. They are separate from the final `logipuz` evaluation bundle: the evaluation bundle contains the selected prompts and task rows, while this package contains the much larger reusable candidate universe.

## Contents

```text
data/pools/<shape>/
  manifest.json
  shards/
  identity_index.jsonl
  payload_index.jsonl
  conflicts/
  difficulty_bins/
```

The release excludes transient `staging/` folders, logs, caches, and locks.

## Summary

- Shapes: `2x2` through `6x6`
- Core candidates: `993,387`
- Difficulty-bin sidecars: included for every shape; production selection enables only nondegenerate shapes by default
- Conflict policy: fixed core structural conflict policy used by LogiPuz selection
- Primary release manifest: `release_manifest.json`
- Integrity checksums: `SHA256SUMS`
- Dataset Viewer preview: `viewer/shape_summary.jsonl`, one row per puzzle shape

## Usage

Place this package at the production data root expected by LogiPuz:

```bash
cp -a data/pools /path/to/logipuz/data/
logipuz pool inspect
logipuz dataset support --count 100
```

The production dataset recipe uses these pools in frozen mode by default. Pool preparation is expected to be current before dataset creation.

## Relationship To `logipuz`

Use `logipuz-pools` when you want to inspect, reuse, or extend the candidate pools. Use the separate `logipuz` dataset artifact when you want the benchmark-ready prompts, task rows, matched controls, bins, audit reports, and evaluation metadata.

The final `logipuz` artifact should reference this pool artifact as provenance, but it should not duplicate the full pools.

## Responsible Use

These are synthetic logic-puzzle candidates. Difficulty bins are solver-derived structural proxies, not validated human difficulty labels. The pools are designed for controlled reasoning benchmark construction and are not intended as general-purpose factual knowledge data.

This dataset is released under CC BY 4.0. Please cite or link the dataset when reusing the pools.