---
  license: apache-2.0
  task_categories:
    - question-answering
  tags:
    - sudoku
    - reasoning
    - hierarchical-reasoning-model
    - puzzle-solving
  pretty_name: Sudoku-Extreme & Augmented Datasets for HRM Reproduction
  ---

  # Sudoku-Extreme Datasets for HRM Reproduction

  Preprocessed datasets used for reproducing the **Hierarchical Reasoning Model (HRM)** and **Augmented HRM** papers on the Sudoku-Extreme benchmark.

  ## Datasets

  | Dataset | Description | Train Examples | Test Examples |
  |---------|-------------|---------------|--------------|
  | `sudoku-extreme-1k-aug-1000` | Vanilla Sudoku-Extreme (1000 puzzles, 1000x augmented) | 1,001,000 | 422,786 |
  | `sudoku-extreme-1k-aug-1000-hint` | Augmented version with easier puzzles mixed in | 2,002,000 | 422,786 |

  ## Source Papers

  - **HRM**: [Hierarchical Reasoning Model](https://arxiv.org/abs/2506.21734) (Wang et al., 2025)
  - **Augmented HRM**: [Are Your Reasoning Models Reasoning or Guessing?](https://arxiv.org/abs/2601.10679) (Ren & Liu, 2026)

  ## Dataset Format

  Each dataset contains `train/` and `test/` directories with:

  - `all__inputs.npy` — Puzzle inputs, shape `(N, 81)`, values 1-10
  - `all__labels.npy` — Solution labels, shape `(N, 81)`, values 2-10
  - `all__puzzle_indices.npy` — Cumulative indices marking puzzle boundaries
  - `all__puzzle_identifiers.npy` — Puzzle type IDs
  - `all__group_indices.npy` — Cumulative group boundary indices
  - `dataset.json` — Metadata (vocab_size=11, seq_len=81, pad_id=0)

  ## How to Build From Scratch

  ```bash
  # Vanilla dataset
  python dataset/build_sudoku_dataset.py --output-dir data/sudoku-extreme-1k-aug-1000 --subsample-size 1000 --num-aug 1000

  # Augmented (hint) dataset
  python dataset/build_sudoku_dataset.py --output-dir data/sudoku-extreme-1k-aug-1000-hint --subsample-size 1000 --num-aug 1000 --hint