emiliocantuc/sudoku-extreme-1k-aug-1000

The sudoku-extreme-1k-aug-1000 used in the Tiny Recursion Models (TRM) paper hosted here for convenience.

NOTE: Value encoding. The original TRM data uses 1..10. We shift everything down by 1, so final values are 0..9 where 0 = empty and 1..9 = digits 1..9.

Generated with the official code by running:

python dataset/build_sudoku_dataset.py --output-dir data/sudoku-extreme-1k-aug-1000  --subsample-size 1000 --num-aug 1000  # 1000 examples, 1000 augments

and then

from pathlib import Path
import numpy as np
from datasets import Dataset, DatasetDict, Features, Sequence, Value

data_dir = Path("./data/sudoku-extreme-1k-aug-1000/")

def load_split(split):
    d = data_dir / split
    # note: shift down by 1 to have values in 0..9 instead of 1..10
    X = np.load(d / "all__inputs.npy") - 1
    y = np.load(d / "all__labels.npy") - 1
    # exclude last entry to match shapes
    idx = np.load(d / "all__puzzle_indices.npy")[:-1]
    # ids = np.load(d / "all__puzzle_identifiers.npy") # omitting since all zeros
    return X, y, idx.astype(np.int32)


X_train, y_train, idx_train = load_split("train")
X_test, y_test, idx_test = load_split("test")

features = Features(
    {
        "inputs": Sequence(Value("uint8"), length=81),
        "labels": Sequence(Value("uint8"), length=81),
        "puzzle_index": Value("int32"),
    }
)

ds_train = Dataset.from_dict(
    {"inputs": X_train, "labels": y_train, "puzzle_index": idx_train}, features=features
)
ds_test = Dataset.from_dict(
    {"inputs": X_test, "labels": y_test, "puzzle_index": idx_test}, features=features
)

dset = DatasetDict({"train": ds_train, "test": ds_test})

repo_id = "emiliocantuc/sudoku-extreme-1k-aug-1000"
dset.push_to_hub(repo_id, private=False)

Example usage in this repo.