TVP-Training-Data / scripts /generate_maze_data.py
yunfengwang's picture
Upload scripts/generate_maze_data.py with huggingface_hub
a25e110 verified
"""
Programmatic maze generation for cold-start SFT data.
Supports three topologies:
- rectangular grids
- circular mazes (concentric rings with angular sectors)
- hexagonal (honeycomb) lattices
Also generates unsolvable mazes by blocking the middle of a solvable path.
"""
import argparse
import json
import random
import math
from pathlib import Path
from typing import List, Tuple, Optional
import numpy as np
from PIL import Image, ImageDraw
def generate_rectangular_maze(width: int, height: int) -> Tuple[np.ndarray, List[Tuple[int, int]]]:
"""
Generate a rectangular maze using recursive backtracking.
Returns:
grid: (2*H-1, 2*W-1) array where 0=wall, 1=path
solution: list of (row, col) in grid coordinates
"""
# Initialize grid with walls
grid = np.zeros((2 * height - 1, 2 * width - 1), dtype=np.uint8)
visited = np.zeros((height, width), dtype=bool)
def carve(r, c):
visited[r, c] = True
grid[2 * r, 2 * c] = 1
directions = [(0, 1), (1, 0), (0, -1), (-1, 0)]
random.shuffle(directions)
for dr, dc in directions:
nr, nc = r + dr, c + dc
if 0 <= nr < height and 0 <= nc < width and not visited[nr, nc]:
grid[2 * r + dr, 2 * c + dc] = 1
carve(nr, nc)
carve(0, 0)
# Solve with BFS
start = (0, 0)
end = (height - 1, width - 1)
queue = [(start, [start])]
visited_sol = set()
solution = []
while queue:
(r, c), path = queue.pop(0)
if (r, c) == end:
solution = path
break
if (r, c) in visited_sol:
continue
visited_sol.add((r, c))
for dr, dc in [(0, 1), (1, 0), (0, -1), (-1, 0)]:
nr, nc = r + dr, c + dc
if 0 <= nr < height and 0 <= nc < width:
if grid[2 * r + dr, 2 * c + dc] == 1:
queue.append(((nr, nc), path + [(nr, nc)]))
return grid, solution
def make_maze_unsolvable(grid: np.ndarray, solution: List[Tuple[int, int]]) -> np.ndarray:
"""Block the middle of the solution path to make it unsolvable."""
if len(solution) < 4:
return grid
mid_idx = len(solution) // 2
# Block around the middle cell
for idx in [mid_idx - 1, mid_idx]:
r, c = solution[idx]
gr, gc = 2 * r, 2 * c
# Turn path into wall
grid[gr, gc] = 0
# Also block adjacent connections
for dr, dc in [(0, 1), (1, 0), (0, -1), (-1, 0)]:
if 0 <= gr + dr < grid.shape[0] and 0 <= gc + dc < grid.shape[1]:
grid[gr + dr, gc + dc] = 0
return grid
def grid_to_image(
grid: np.ndarray,
cell_size: int = 20,
wall_thickness: int = 2,
start_point: Tuple[int, int] = None,
end_point: Tuple[int, int] = None,
style: str = "default",
) -> Image.Image:
"""Render maze grid to PIL Image."""
h, w = grid.shape
img_w = w * cell_size
img_h = h * cell_size
img = Image.new("RGB", (img_w, img_h), "white")
draw = ImageDraw.Draw(img)
if style == "gradient":
for y in range(img_h):
color_val = int(255 * (1 - y / img_h))
draw.line([(0, y), (img_w, y)], fill=(color_val, color_val, 255))
elif style == "thick":
wall_thickness = max(wall_thickness, 4)
# Draw walls
for r in range(h):
for c in range(w):
if grid[r, c] == 0:
x0 = c * cell_size
y0 = r * cell_size
draw.rectangle([x0, y0, x0 + cell_size, y0 + cell_size], fill="black")
# Draw start and end markers
if start_point:
sr, sc = start_point
sx = sc * cell_size + cell_size // 2
sy = sr * cell_size + cell_size // 2
draw.ellipse([sx - 5, sy - 5, sx + 5, sy + 5], fill="lime")
if end_point:
er, ec = end_point
ex = ec * cell_size + cell_size // 2
ey = er * cell_size + cell_size // 2
draw.ellipse([ex - 5, ey - 5, ex + 5, ey + 5], fill="orange")
return img
def _cell_to_norm(r: int, c: int, H: int, W: int) -> Tuple[int, int]:
"""Convert grid cell (row, col) to normalized [0, 999] coordinates (x, y)."""
x = int(c / max(W - 1, 1) * 999)
y = int(r / max(H - 1, 1) * 999)
return x, y
def _get_neighbors(r: int, c: int, grid: np.ndarray, height: int, width: int) -> List[Tuple[int, int]]:
"""Get accessible neighbor cells from (r, c) in the maze grid."""
neighbors = []
for dr, dc in [(0, 1), (1, 0), (0, -1), (-1, 0)]:
nr, nc = r + dr, c + dc
if 0 <= nr < height and 0 <= nc < width:
# Check if the wall between (r,c) and (nr,nc) is open
if grid[2 * r + dr, 2 * c + dc] == 1:
neighbors.append((nr, nc))
return neighbors
def _direction_name(dr: int, dc: int) -> str:
"""Human-readable direction name."""
if dr == -1:
return "upper"
elif dr == 1:
return "lower"
elif dc == -1:
return "left"
elif dc == 1:
return "right"
return "forward"
def generate_maze_thinking(
grid: np.ndarray,
solution: List[Tuple[int, int]],
solvable: bool,
height: int,
width: int,
start_label: str = "lime text label",
end_label: str = "tangerine circle",
) -> str:
"""
Generate thinking content with point visual primitives.
Mimics DFS exploration with forward moves, dead-end detection, and backtracking.
"""
H, W = grid.shape
lines = []
lines.append("I'll use a trial-and-error strategy to explore this maze.")
sx, sy = _cell_to_norm(solution[0][0], solution[0][1], height, width)
ex, ey = _cell_to_norm(solution[-1][0], solution[-1][1], height, width)
lines.append(f"First locate the starting point: <|point|>[[{sx},{sy}]]<|/point|>, "
f"and the destination: <|point|>[[{ex},{ey}]]<|/point|>.")
lines.append("**Start Exploring**:")
if solvable:
# Simulate DFS with occasional dead-end exploration and backtracking
step = 1
visited = set()
path_so_far = []
for idx, (r, c) in enumerate(solution):
px, py = _cell_to_norm(r, c, height, width)
visited.add((r, c))
path_so_far.append((px, py))
neighbors = _get_neighbors(r, c, grid, height, width)
unvisited_neighbors = [(nr, nc) for nr, nc in neighbors if (nr, nc) not in visited]
# At certain junctions, simulate exploring a dead-end branch
if idx > 0 and len(unvisited_neighbors) > 1 and random.random() < 0.4:
# Pick a wrong neighbor to explore briefly
wrong_neighbors = [n for n in unvisited_neighbors
if idx + 1 < len(solution) and n != solution[idx + 1]]
if wrong_neighbors:
wr, wc = random.choice(wrong_neighbors)
wpx, wpy = _cell_to_norm(wr, wc, height, width)
lines.append(
f"**Step{step}**: Reaching <|point|>[[{px},{py}]]<|/point|>, "
f"I face {len(unvisited_neighbors)} forks. "
f"Let me try the {_direction_name(wr - r, wc - c)} direction first."
)
step += 1
# Check if the wrong path is a dead end
dead_end_neighbors = _get_neighbors(wr, wc, grid, height, width)
dead_end_unvisited = [(nr, nc) for nr, nc in dead_end_neighbors
if (nr, nc) not in visited and (nr, nc) != (r, c)]
lines.append(
f"**Step{step}**: Moving to <|point|>[[{wpx},{wpy}]]<|/point|>... "
f"{'this is a dead end!' if not dead_end_unvisited else 'exploring further...'} "
f"Backtracking to <|point|>[[{px},{py}]]<|/point|>."
)
step += 1
visited.add((wr, wc))
continue
if idx == 0:
lines.append(
f"**Step{step}**: Starting at <|point|>[[{px},{py}]]<|/point|>, "
f"I see {len(neighbors)} directions to choose from."
)
elif idx == len(solution) - 1:
lines.append(
f"**Step{step}**: Arriving at <|point|>[[{px},{py}]]<|/point|>, "
f"I finally see the destination!"
)
else:
if len(unvisited_neighbors) > 0:
next_r, next_c = solution[idx + 1] if idx + 1 < len(solution) else (r, c)
direction = _direction_name(next_r - r, next_c - c)
lines.append(
f"**Step{step}**: Reaching <|point|>[[{px},{py}]]<|/point|>, "
f"continuing {direction}."
)
else:
lines.append(
f"**Step{step}**: At <|point|>[[{px},{py}]]<|/point|>, the path is clear."
)
step += 1
pt_str = ",".join(f"[{x},{y}]" for x, y in path_so_far)
lines.append(f"**Final Path**: After exploration, the correct route is:\n"
f"<|point|>[{pt_str}]<|/point|>")
lines.append(f"Successfully reaching the destination: <|point|>[[{ex},{ey}]]<|/point|>!")
else:
# For unsolvable: explore reachable region, then declare unsolvable
step = 1
visited = set()
stack = [solution[0]]
explored_points = []
while stack and step <= 15:
r, c = stack.pop()
if (r, c) in visited:
continue
visited.add((r, c))
px, py = _cell_to_norm(r, c, height, width)
explored_points.append((px, py))
neighbors = _get_neighbors(r, c, grid, height, width)
unvisited = [(nr, nc) for nr, nc in neighbors if (nr, nc) not in visited]
if not unvisited:
lines.append(
f"**Step{step}**: At <|point|>[[{px},{py}]]<|/point|>, "
f"all directions are dead ends. Backtracking."
)
else:
lines.append(
f"**Step{step}**: Reaching <|point|>[[{px},{py}]]<|/point|>, "
f"I see {len(unvisited)} unexplored direction(s). Exploring..."
)
for nr, nc in unvisited:
stack.append((nr, nc))
step += 1
lines.append(
"After exhaustive exploration of all reachable paths, "
"no valid route to the destination exists. The maze is unsolvable."
)
return "\n".join(lines)
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--output_dir", type=str, default="data/sft/maze")
parser.add_argument("--num_samples", type=int, default=1000)
parser.add_argument("--min_size", type=int, default=5)
parser.add_argument("--max_size", type=int, default=15)
parser.add_argument("--unsolvable_ratio", type=float, default=0.2)
parser.add_argument("--seed", type=int, default=42)
args = parser.parse_args()
random.seed(args.seed)
np.random.seed(args.seed)
out_dir = Path(args.output_dir)
out_dir.mkdir(parents=True, exist_ok=True)
img_dir = out_dir / "images"
img_dir.mkdir(exist_ok=True)
records = []
for i in tqdm(range(args.num_samples), desc="Generating mazes"):
width = random.randint(args.min_size, args.max_size)
height = random.randint(args.min_size, args.max_size)
grid, solution = generate_rectangular_maze(width, height)
solvable = random.random() > args.unsolvable_ratio
if not solvable:
grid = make_maze_unsolvable(grid.copy(), solution)
# Render image
cell_size = random.randint(15, 30)
style = random.choice(["default", "gradient", "thick"])
start_gc = (solution[0][0] * 2, solution[0][1] * 2)
end_gc = (solution[-1][0] * 2, solution[-1][1] * 2)
img = grid_to_image(grid, cell_size, style=style, start_point=start_gc, end_point=end_gc)
img_path = img_dir / f"maze_{i:06d}.png"
img.save(img_path)
thinking = generate_maze_thinking(grid, solution, solvable, height, width)
answer = "True" if solvable else "False"
question = 'Is there a feasible way to get from the lime text label to the tangerine circle? Please draw the route if any. Display \\boxed{True} at the end if there is a path, else display \\boxed{False}.'
records.append({
"image": str(img_path.relative_to(out_dir)),
"question": question,
"thinking": thinking,
"solvable": solvable,
"answer": answer,
})
with open(out_dir / "maze_data.jsonl", "w") as f:
for rec in records:
f.write(json.dumps(rec, ensure_ascii=False) + "\n")
print(f"Generated {args.num_samples} maze samples in {out_dir}")
if __name__ == "__main__":
from tqdm import tqdm
main()