| --- |
| license: mit |
| tags: |
| - puzzles |
| - reasoning |
| - pencil-puzzles |
| viewer: false |
| --- |
| |
| # ICS datasets |
|
|
| The Light-Up, Nurikabe, Tapa and Heyawake datasets of ICS (code: https://github.com/icsearch/ICS). Each folder holds a |
| `test` and a `train` split in the layout the code reads: |
|
|
| | Folder | Puzzle | Test boards (random + golden) | Train boards | Train rows | |
| |---|---|---|---|---| |
| | `lightup-ms26` | Light-Up | 113 (100 + 13) | 1,431 | 11,448 | |
| | `ppb-nurikabe-ms26` | Nurikabe | 115 (100 + 15) | 2,282 | 22,820 | |
| | `ppb-tapa-ms26` | Tapa | 65 (50 + 15) | 580 | 5,800 | |
| | `ppb-heyawake-ms26` | Heyawake | 115 (100 + 15) | 5,564 | 55,640 | |
|
|
| ## Loading |
|
|
| ```python |
| from huggingface_hub import snapshot_download |
| |
| root = snapshot_download(repo_id="hcshi/ICS-data", repo_type="dataset", allow_patterns="ppb-tapa-ms26/*") |
| ``` |
|
|
| With the code installed, `python -m ics download --dataset --task tapa` writes the same folder to `data/tapa`. Pass |
| a folder, e.g. `<root>/ppb-tapa-ms26`, as `--data` to `python -m ics train` or `python -m ics eval`. |
|
|
| ## Source and license |
|
|
| The puzzles come from Pencil Puzzle Bench (PPBench), the dataset `bluecoconut/pencil-puzzle-bench` at revision |
| `3ac6add0754f1458bf51aab160ca039667a67b38`, under the MIT License: see `LICENSE-PPBench.txt`. |
|
|
