README.md

---
pretty_name: FinalSudoku
license: odc-by
task_categories:
- other
language:
- en
tags:
- sudoku
- benchmark
- reasoning
- verifiable-reasoning
- exact-match
- puzzle-solving
- logic
- constraint-satisfaction
- combinatorial-search
- model-evaluation
size_categories:
- 1K<n<10K
configs:
- config_name: default
  data_files:
  - split: test
    path: data/public/test-*.jsonl
---

# FinalSudoku

**FinalSudoku** is a Sudoku benchmark for evaluating exact, verifiable puzzle-solving in AI systems.

The public release contains 1000 extremely hard Sudoku puzzles, averaging an **ER of 11.114** with verified answers. We also maintain a separate **3000-puzzle private evaluation set** for official leaderboard results.

## Format

Each puzzle is represented as an 81-character string in row-major order.

- Digits `1`–`9` are givens.
- `.` means an empty cell.
- The answer is the completed 81-digit solution string.

Example:

```json
{
    "id": "FS-PUBLIC-TEST-000018", 
    "question": ".....1..2....3..4...56..7....6..3....5.8..9..9...1..6..67.....4.89.2....5..7..8..", 
    "answer": "648571392791238645235694781826953417154867923973412568367189254489325176512746839", 
    "er": 11.0, 
    "ep": 11.0, 
    "ed": 3.4
}
```

## Fields

| Field | Description |
|---|---|
| `id` | Stable puzzle ID |
| `question` | 81-character Sudoku puzzle, with `.` for empty cells |
| `answer` | 81-character solved Sudoku grid |
| `er` | Explainer Rating |
| `ep` | Pearl Rating |
| `ed` | Diamond Rating |

## Methodology

We computed ER/EP/ED difficulty ratings with SukakuExplainer’s `serate` solver. Each puzzle was converted to 81-character Sudoku grid format and rated using SukakuExplainer.jar with the legacy-compatible `--format=%i ED=%r/%p/%d` output. The published ratings are the raw SukakuExplainer outputs, with no manual adjustment.

According to these calculations, the average rating for FinalSudoku is 11.114 ER. 

## Usage

```python
from datasets import load_dataset

ds = load_dataset("faranglab/FinalSudoku", split="test")
print(ds[0])
```

## Evaluation

The main metric is **exact solve rate**.

A response is correct only if the final parsed answer:

1. Is an 81-digit string.
2. Forms a valid completed Sudoku grid.
3. Respects all givens in the `question`.
4. Exactly matches the `answer`.

Example LLM prompt:

```text
You are a careful Sudoku solver. The user gives one 9x9 Sudoku puzzle as a flat 81-character row-major string. Digits 1-9 are fixed clues and '.' means blank. Your job is to produce a complete valid Sudoku solution that preserves every given clue. Before answering, internally verify all of the following: (1) every original clue is unchanged, (2) every row contains digits 1-9 exactly once, (3) every column contains digits 1-9 exactly once, and (4) every 3x3 box contains digits 1-9 exactly once. If any check fails, keep fixing the board internally until all checks pass. Return only the final checked answer in the required JSON format.

Solve this Sudoku carefully.

Puzzle (row-major 81 chars, '.' means blank):
{puzzle}

Requirements before you answer:
- preserve every given clue exactly
- make every row valid
- make every column valid
- make every 3x3 box valid
- if your first internal attempt fails any check, fix it before answering

Return JSON only in this exact shape:
{{\"answer\":\"<81 digits>\"}}
Do not include markdown fences, explanations, reasoning, or extra keys.
```
## Reproducing the results

Eval scripts are available at [Github-repository](https://github.com/FarangLab/FinalSudoku)

## Leaderboard

To be published

## Licensing Information

This collection is released under the Open Data Commons Attribution License (ODC-By) v1.0 [license](https://opendatacommons.org/licenses/by/1-0/).

## Citation

```bibtex
@misc{finalsudoku2026,
  title = {FinalSudoku: A Benchmark for Extremely Hard Sudoku Reasoning},
  author = {Salam Jadari},
  year = {2026},
  url = {https://huggingface.co/datasets/faranglab/FinalSudoku}
}
```