selimaktas/maze-curriculum-dataset

Maze Navigation Curriculum Dataset

Dataset Description

This dataset contains 60,000 ASCII maze puzzles designed for training language models on spatial reasoning and navigation tasks. The dataset implements curriculum learning with progressively larger mazes.

Dataset Structure

Data Instances

Each instance contains:

• prompt: Instruction for the navigation task
• maze: ASCII representation of the maze with '#' for walls, 'S' for start, 'E' for end
• solution: Sequence of moves (N/S/E/W) to solve the maze
• grid_size: Navigable path size (5, 7, or 9)

Data Fields

• prompt (string): "You are a navigation agent. Given an ASCII maze, move from the start cell 'S' to the end cell 'E'. Respond with a sequence of moves using the letters N, S, W, E"
• maze (string): ASCII maze representation
• solution (string): Solution path as sequence of directional moves
• grid_size (int): Size of the navigable area

Data Splits

• train: 60,000 examples
  • 5x5 mazes: 25,000 examples
  • 7x7 mazes: 20,000 examples
  • 9x9 mazes: 15,000 examples

Example

Maze:
#########
#S      #
# ##### #
#     # #
##### # #
#     # #
# ####### 
#       E#
#########

Solution: EEEESSSSWWWWSSEEEEE

Usage

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("selimaktas/maze-curriculum-dataset")

# Access training data
train_data = dataset["train"]

# Example usage
for example in train_data:
    print("Maze:")
    print(example["maze"])
    print("Solution:", example["solution"])
    break

Curriculum Learning

The dataset is designed for curriculum learning:

1. Start with 5x5 mazes (simpler)
2. Progress to 7x7 mazes (medium complexity)
3. Advance to 9x9 mazes (more challenging)

Generation Details

• Generated using Wilson's algorithm for uniform maze distribution
• Each maze has exactly one solution path
• ASCII format with '#' for walls, ' ' for paths, 'S' for start, 'E' for end
• Navigable grid sizes refer to open path areas (actual grid is larger due to walls)

Citation

If you use this dataset, please cite it appropriately in your research.