twistshan/Sudoku_DLM_Reasoning

Sudoku DLM Reasoning

stwistzz/Sudoku_DLM_Reasoning is a deterministic 9x9 Sudoku benchmark for studying masked diffusion language models, depth, and iterative decoding. Version 1.1.0 trains only on original, r0, and r1_4; r5_19 is held out for adjacent difficulty extrapolation.

Version 1.1.0 changes only the deterministic selection seed to 0 relative to v1.0.2. The raw CSV files, sources, bucket definitions, split quotas, schema, and validation rules are unchanged; the four Parquet splits are regenerated deterministically from the same candidate pool.

Repository layout

• raw/: complete 1,415,420-row candidate snapshot used by this protocol
• data/: selected, leakage-audited Parquet splits used for experiments
• metadata/: source/release hashes, split specification, and audits
• scripts/: deterministic builder and independent verifier
• docs/BUILDING.md: end-to-end reconstruction instructions
• docs/DIFFICULTY.md: operational difficulty definition and quantitative audit

Files under raw/ are explicitly source data and are not loaded as extra splits. Only the four data/*.parquet paths declared in the card metadata form the default dataset config.

Splits

split	original	r0	r1_4	r5_19	total
train	65,536	65,536	65,536	0	196,608
validation	2,048	2,048	2,048	0	6,144
test	1,000	1,000	1,000	1,000	4,000
test_ood_confirm	0	0	0	4,000	4,000

The primary zero-shot protocol selects the checkpoint, decoding strategy, and number of decoding steps using only validation. It must not use r5_19 before the final test evaluation. test_ood_confirm is reserved for confirming a small number of already-selected configurations; it is not a tuning split.

Difficulty

r0, r1_4, and r5_19 use an upstream Tdoku search measurement. In public Tdoku commit af426180dc53aef89b82868e7b3fdfcf42165654, the counter increments when the solver expands a binary band/stack-configuration decision after constraint propagation. It is therefore best interpreted as visited search decision nodes, not recursion depth, failed undo count, clue count, or human difficulty. original comes from a separate easy collection and has no official rating on this scale.

See docs/DIFFICULTY.md for the DFS root, propagation and branch semantics, bucket formulas, full raw-pool statistics, source confounding, and the upstream rating-provenance limitation.

The evaluation sample deliberately covers subranges:

• r1_4: [1,2), [2,3), [3,4), [4,5)
• r5_19: [5,8), [8,12), [12,16), [16,20)

The 1,000-row test allocation uses 250 rows from each subrange. The 4,000-row confirmation allocation uses 1,000 rows from each r5_19 subrange. Within a subrange, source collection and clue count retain their upstream proportions.

Deterministic construction

• Upstream repository: fhyfhy/diffusion-vs-ar-hard-sudoku
• Pinned upstream revision: 527859f62c745c16833aded130ad9f9ddddb76af
• Release seed: 0
• Sampling: proportional metadata-stratified bottom-k by SHA-256 rank
• Leakage checks: exact puzzle plus digit-renaming-normalized puzzle/solution pair
• Validation: 81-character format, clue consistency, and Sudoku row/column/box validity

The source train/test boundary is preserved. Validation rows come only from unused upstream training rows. Test and confirmation rows come only from upstream test files. Full file hashes, row provenance, exclusions, and pairwise overlap checks are in metadata/manifest.json and metadata/audit.json.

The upstream Sudoku Extreme corpus states that its official train and test sets are mathematically inequivalent. This release additionally audits exact and digit renaming overlap across the mixed original/Extreme sources. It does not implement full canonicalization over every Sudoku row, column, band, stack, and transpose symmetry; that remains a documented limitation.

Schema

The main columns are puzzle, solution, difficulty_bucket, difficulty_subbucket, official_rating, clues, source_collection, and release_split. example_id, puzzle_id, and digit_normalized_id are stable SHA-256 identifiers. source_file and source_row_index provide exact provenance.

from datasets import load_dataset

dataset = load_dataset("stwistzz/Sudoku_DLM_Reasoning")
train = dataset["train"]
validation = dataset["validation"]
test = dataset["test"]

Puzzles and solutions are 81-character row-major strings. 0 denotes an empty cell in puzzle.

Rebuild

The pinned source CSV files are included under raw/. Run:

python scripts/build_dataset.py \
  --source-dir raw \
  --output-dir /path/to/release \
  --dataset-id stwistzz/Sudoku_DLM_Reasoning \
  --seed 0

The builder requires Python 3.10+ and PyArrow. See docs/BUILDING.md, then run scripts/verify_dataset.py against the rebuilt release.

Attribution and license

The data is derived from fhyfhy/diffusion-vs-ar-hard-sudoku, which packages the original Sudoku data from HKUNLP/diffusion-vs-ar and the sapientinc/sudoku-extreme corpus. Please consult and comply with the licenses and attribution requirements of all upstream sources. This derived release does not grant rights beyond them. See RAW_DATA_LICENSES.md for the raw-snapshot notice.