erdem-erdem/maze-llm-dataset

Multi-turn Maze Navigation Dataset

A dataset of maze navigation sequences for training AI agents to solve mazes through sequential decision-making.

Dataset Description

This dataset contains unique mazes with complete solution sequences. Each example includes:

• —A maze with visual path tracking
• —Sequential moves from start to goal
• —Minimum 2 moves per sequence
• —At least 2 different directions required per maze

Dataset Stats

Train Set

Total Sequences: 100,836 Total Unique Mazes: 100,836 Total Moves: 1,564,964 Avg Moves per Maze: 15.5

Curriculum Distribution:

3x3: 5,836 mazes (5.8%), 41,624 moves 4x4: 19,000 mazes (18.8%), 187,556 moves 5x5: 19,000 mazes (18.8%), 242,408 moves 6x6: 19,000 mazes (18.8%), 302,866 moves 7x7: 19,000 mazes (18.8%), 362,866 moves 8x8: 19,000 mazes (18.8%), 427,644 moves

Quality Analysis:

Action distribution: N: 393,144, S: 389,608, E: 390,008, W: 392,204 Sequence length range: 4-76 Avg sequence length: 15.5

Test Set

Total Sequences: 5,600 Total Unique Mazes: 5,600 Total Moves: 83,478 Avg Moves per Maze: 14.9

Curriculum Distribution:

3x3: 600 mazes (10.7%), 4,234 moves 4x4: 1,000 mazes (17.9%), 9,752 moves 5x5: 1,000 mazes (17.9%), 12,730 moves 6x6: 1,000 mazes (17.9%), 15,878 moves 7x7: 1,000 mazes (17.9%), 18,878 moves 8x8: 1,000 mazes (17.9%), 22,006 moves

Quality Analysis:

Action distribution: N: 21,142, S: 20,682, E: 21,266, W: 20,388 Sequence length range: 4-62 Avg sequence length: 14.9

Data Format

Each example contains:

• —prompt: Navigation instruction
• —legend: Symbol definitions
• —sequence: List of turns with maze states and correct moves
• —total_steps: Number of moves in sequence
• —grid_size: Maze dimensions
• —full_solution: Complete solution string
• —maze_id: Unique identifier