emiliocantuc/maze-30x30-hard-1k

The maze-30x30-hard-1k used in the Tiny Recursion Models (TRM) paper hosted here for convenience.

NOTE: Value encoding. The original TRM data uses 1..5. We shift everything down by 1, so final values are 0..4 The mapping is:

• # wall -> 0
• space/open cell -> 1
• S start -> 2
• G goal -> 3
• o solution path -> 4

Use seq_len=900 and vocab_size=5 with this dataset.

Generated with the official code by running:

python dataset/build_maze_dataset.py --aug # 1000 examples, 8 augments

and then

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

data_dir = Path("./data/maze-30x30-hard-1k/")

def load_split(split):
    d = data_dir / split
    # note: shift down by 1 to have values in 0..4 instead of 1..5
    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=30 * 30),
        "labels": Sequence(ClassLabel(names=["#", " ", "S", "G", "o"]), length=30 * 30),
        "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/maze-30x30-hard-1k"
dset.push_to_hub(repo_id, private=False)

Example usage in this repo.