SudokuDiT

A compact Diffusion Transformer (1.28 M params) that solves Sudoku as a masked discrete diffusion (MDLM-style) denoiser: it fills a puzzle by iteratively un-masking the cells it is most confident about, MaskGIT-style.

watch it solve

  • Architecture: DiT with adaLN-Zero conditioning β€” hidden=128, heads=4, blocks=4; per-cell token + 2-D positional + 3x3-box embeddings, plus a timestep.
  • Code, training and an interactive web demo: https://github.com/tchauffi/nonet

Input / output

A board is 81 tokens, row-major: 0 = empty / [MASK], 1..9 = digits. The solver clamps the given clues and only fills the blanks.

Usage

import torch
from nonet.hub import load_solver   # pip install git+https://github.com/tchauffi/nonet

solver = load_solver("tchauffi/sudoku-dit")     # downloads model.safetensors + config.json
puzzle = "402609001000000000700002084000080500030005007004060930008021406005000000147030000"
x = torch.tensor([[int(c) for c in puzzle]])      # (1, 81), 53 blanks
solution = solver.solve(x, conf_threshold=0.999)  # adaptive reveal (recommended)
# fixed-budget alternative: solver.solve(x, num_steps=81)

"".join(map(str, solution[0].tolist()))
# 482659371613874295759312684971483562836295147524167938398721456265948713147536829

Performance

Held-out validation puzzles, adaptive decoder (tau = 0.999):

metric value
valid solutions 99.8 %
exact match vs reference 94.3 %
avg reveal steps / puzzle 3.6

Solve rate is 100 % up to ~39 blanks and ~98 % on the hard 50+ blank tail (where the adaptive decoder spends more steps). exact_match is lower than valid only because the dataset's high-blank puzzles aren't always uniquely solvable, so the model may return a different valid grid.

Training

  • Data: Ritvik19/Sudoku-Dataset (~17 M puzzles), tokenized on the fly.
  • Objective: masked cross-entropy over masked cells, conditional (given clues are never masked), with a linear masking schedule.
  • ~50 k steps, AdamW, batch 256. Kept deliberately small β€” larger models sit at the entropy floor far longer before the loss breaks through.

Limitations

  • Not a guaranteed solver: a small fraction of very hard (60+ blank) boards come out invalid.
  • Trained only on standard 9x9 Sudoku from the dataset above.
Downloads last month
15
Safetensors
Model size
1.28M params
Tensor type
F32
Β·
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Dataset used to train tchauffi/sudoku-dit

Collection including tchauffi/sudoku-dit