PotatoAGI (RHAN-Sudoku)

This is the official weight repository for the Recurrent Hybrid Attention Network (RHAN) trained on Sudoku.

It uses a Universal Linear Attention mechanism combined with Recursive Memory and was trained using Adversarial Erasure.

Stats

• Parameters: ~150k
• Architecture: 12-Loop Recurrent CNN + Linear Attention
• Accuracy: 99% Cell Accuracy / 90%+ Perfect Solve Rate
• License: CC BY-NC 4.0 (Non-Commercial Research Use Only)

Files in this Repository

model.py # Model architecture (UniversalPotato) model.safetensors # Trained weights local_test_sudoku.py # Dataset-based local evaluation README.md

Usage

1️⃣ Install dependencies

pip install torch safetensors

Python ≥ 3.10 recommended.

2️⃣ Load the model and weights

import torch from safetensors.torch import load_file from model import UniversalPotato, HIDDEN_DIM

device = "cuda" if torch.cuda.is_available() else "cpu"

model = UniversalPotato().to(device) model.load_state_dict(load_file("model.safetensors"), strict=True) model.eval()

3️⃣ Run inference on a single Sudoku puzzle

Sudoku grids are represented as a flat tensor of length 81, with 0 indicating empty cells.

Example puzzle (0 = empty)

puzzle = [ 5,3,0,0,7,0,0,0,0, 6,0,0,1,9,5,0,0,0, 0,9,8,0,0,0,0,6,0, 8,0,0,0,6,0,0,0,3, 4,0,0,8,0,3,0,0,1, 7,0,0,0,2,0,0,0,6, 0,6,0,0,0,0,2,8,0, 0,0,0,4,1,9,0,0,5, 0,0,0,0,8,0,0,7,9, ]

clues = torch.tensor(puzzle, dtype=torch.long).unsqueeze(0).to(device) board = clues.clone() memory = torch.zeros(1, HIDDEN_DIM, 9, 9, device=device)

with torch.no_grad(): for _ in range(24): # reasoning steps logits, memory = model( clues=clues, current_board=board, memory=memory, blindfold=False, ) board = logits.argmax(dim=-1)

solution = board.view(9, 9).cpu() print(solution)

4️⃣ Dataset-based evaluation

To evaluate the model on a real Sudoku dataset:

Download sudoku.csv from Kaggle
👉 https://www.kaggle.com/datasets/rohanrao/sudoku

Place it in the repository root

Run:

python local_test_sudoku.py

This script:

runs multi-step inference

compares predictions against ground truth

reports solve success rate

Notes

This model does not use Hugging Face Transformers

model.py is the authoritative architecture definition

Inference requires multiple recurrent steps for best results

Designed for reasoning research, not commercial deployment

License

This project is released under CC BY-NC 4.0.

You may:

use

modify

redistribute
for non-commercial research purposes only, with attribution.

Commercial use is not permitted.