Show HN: Training a sudoku solver from scratch on Jetson Nano

A developer released a from-scratch Sudoku solver built on a Looped MLP-Mixer architecture, trained on the sapientinc/sudoku-extreme dataset and runnable on an NVIDIA Jetson Nano via JetPack Docker. The model reuses a shared MLP-Mixer stack across an outer commit loop with inner mixer iterations and a learned halt head, and ships with download-dataset, train, eval, and resume entry points under an MIT License. Each run writes history.json, checkpoints, and per-puzzle trajectory JSON to runs/<run-id>/, with a local viz server charting train/val loss, cell and puzzle accuracy, halt rate, and accuracy by rating group.

Looped MLP-Mixer sudoku solver with an outer commit loop, inner mixer iterations, and a learned halt head, trained from scratch on sapientinc/sudoku-extreme https://huggingface.co/datasets/sapientinc/sudoku-extreme . Install with uv sync , download the dataset ~798 MB , train, and open the local viz server. Minimal path: uv sync uv run download-dataset uv run train \ --min-rating 0 --max-rating 0 \ --max-samples 100 --epochs 30 \ --dim 512 --num-blocks 2 \ --inner-iters 5 --train-max-outer-iters 10 \ --eval-max-outer-iters 10 \ --train-batch-size 8 --batches-per-epoch 100 uv run python -m http.server 8000 Open http://localhost:8000/viz/ http://localhost:8000/viz/ . Serve from the repo root . Shared MLP-Mixer stack reused across an outer commit loop: each outer step applies the previous prediction, runs inner mixer iterations on encoded grid state, and updates detached cell memory. A halt head learns when the grid matches the solution. Training uses parallel puzzle slots with optional augmentations and partial ground-truth reveal. Entry points: download-dataset , train , eval , resume . Eval supports test-time restarts and one-dimensional compute sweeps inner steps, outer commits, tries . PyTorch profiler hooks are available on train . Each run writes history.json , checkpoints, and per-puzzle trajectory JSON under runs/<run-id / . The viz page charts train/val metrics and plays back outer-commit trajectories. Train and validation metrics per epoch: loss, cell/puzzle accuracy, halt rate, and accuracy by rating group. Trajectory player for one puzzle: model input and output at each outer commit until halt or max steps. Docker setup for training and eval on NVIDIA Jetson JetPack . Source is bind-mounted; rebuild only when dependencies change. MIT License.