Eran92/pot-sudoku-78

PoT Sudoku 78.9% - Pre-trained Checkpoint

This repository contains a pre-trained HybridPoHHRMSolver model achieving 78.9% grid accuracy on Sudoku-Extreme.

Model Details

Metric	Value
Grid Accuracy	78.9%
Cell Accuracy	97.8%
Parameters	20,799,516
Architecture	HybridPoHHRMSolver
Controller	CausalDepthTransformerRouter

Configuration

config = {
    "d_model": 512,
    "n_heads": 8,
    "H_layers": 2,
    "L_layers": 2,
    "d_ff": 2048,
    "H_cycles": 2,
    "L_cycles": 6,
    "T": 4,
    "dropout": 0.039,
    "hrm_grad_style": True,
    "halt_max_steps": 2,
    "controller_type": "transformer",
    "controller_kwargs": {
        "n_ctrl_layers": 2,
        "n_ctrl_heads": 4,
        "d_ctrl": 256,
        "max_depth": 32,
        "token_conditioned": True,
    },
    "injection_mode": "none",
    "vocab_size": 10,
    "num_puzzles": 1,
    "puzzle_emb_dim": 512,
}

Usage

import torch
from huggingface_hub import hf_hub_download
from src.pot.models.sudoku_solver import HybridPoHHRMSolver

# Download checkpoint
checkpoint_path = hf_hub_download("Eran92/pot-sudoku-78", "best_model.pt")

# Create model with exact config
model = HybridPoHHRMSolver(
    d_model=512,
    n_heads=8,
    H_layers=2,
    L_layers=2,
    d_ff=2048,
    H_cycles=2,
    L_cycles=6,
    hrm_grad_style=True,
    halt_max_steps=2,
    controller_type="transformer",
    controller_kwargs={
        "n_ctrl_layers": 2,
        "n_ctrl_heads": 4,
        "d_ctrl": 256,
        "max_depth": 32,
        "token_conditioned": True,
    },
    injection_mode="none",
)

# Load weights
checkpoint = torch.load(checkpoint_path, map_location="cpu")
model.load_state_dict(checkpoint["model_state_dict"])
model.eval()

# Solve a puzzle
puzzle = torch.tensor([[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]])
solution = model.solve(puzzle)
print(solution.reshape(9, 9))

Training

Trained on Sudoku-Extreme (10k puzzles) with:

• Batch size: 768
• Learning rate: 3.7e-4
• Epochs: 2001
• On-the-fly augmentation (digit permutation, transpose, shuffling)

Links

• PoT Repository
• W&B Run