mirror of
https://github.com/open-thought/reasoning-gym.git
synced 2026-04-19 12:58:07 +00:00
7475a20700
* update difficulty metadata for logic datasets * update difficulty metadata for graph datasets * update difficulty metadata for geometry datasets * update difficulty metadata for games datasets * update difficulty metadata for cognition datasets * update difficulty metadata for arithmetic datasets * update difficulty metadata for arc datasets * update difficulty metadata for algorithmic datasets * update difficulty metadata for algebra datasets * use tuples * update tests * update tests
279 lines
9.7 KiB
Python
279 lines
9.7 KiB
Python
"""Sudoku puzzle generator"""
|
|
|
|
import copy
|
|
from dataclasses import dataclass
|
|
from random import Random
|
|
from typing import Any, Optional
|
|
|
|
from ..coaching import BaseCurriculum, RangeAttributeDefinition
|
|
from ..factory import ProceduralDataset, register_dataset
|
|
|
|
|
|
@dataclass
|
|
class SudokuConfig:
|
|
"""
|
|
Configuration for sudoku puzzle generation
|
|
Puzzle generation can be a bit slower for puzzles with a high (~60+) number of empty cells
|
|
"""
|
|
|
|
min_empty: int = 30 # Minimum number of empty cells
|
|
max_empty: int = 50 # Maximum number of empty cells
|
|
seed: Optional[int] = None
|
|
size: int = 500 # Virtual dataset size
|
|
|
|
def validate(self):
|
|
"""Validate configuration parameters"""
|
|
# 81 - 64 = 17, the minimum number of clues required for 9x9 Sudoku to have a unique solution
|
|
assert 0 <= self.min_empty <= 64, "min_empty must be between 0 and 64"
|
|
assert self.min_empty <= self.max_empty <= 64, "max_empty must be between min_empty and 64"
|
|
|
|
|
|
class SudokuDataset(ProceduralDataset):
|
|
"""Generates sudoku puzzles with configurable difficulty"""
|
|
|
|
def __init__(self, config: SudokuConfig):
|
|
super().__init__(config=config, seed=config.seed, size=config.size)
|
|
|
|
def __len__(self) -> int:
|
|
return self.config.size
|
|
|
|
def __iter__(self):
|
|
self._current_idx = 0
|
|
return self
|
|
|
|
def __next__(self):
|
|
if self._current_idx >= self.config.size:
|
|
raise StopIteration
|
|
item = self[self._current_idx]
|
|
self._current_idx += 1
|
|
return item
|
|
|
|
def _is_valid(self, board: list[list[int]], row: int, col: int, num: int) -> bool:
|
|
"""Check if number can be placed at position"""
|
|
# Check row
|
|
if num in board[row]:
|
|
return False
|
|
|
|
# Check column
|
|
if num in [board[i][col] for i in range(9)]:
|
|
return False
|
|
|
|
# Check 3x3 box
|
|
box_row, box_col = 3 * (row // 3), 3 * (col // 3)
|
|
for i in range(box_row, box_row + 3):
|
|
for j in range(box_col, box_col + 3):
|
|
if board[i][j] == num:
|
|
return False
|
|
return True
|
|
|
|
def _get_possible_values(self, board: list[list[int]], row: int, col: int) -> set[int]:
|
|
"""Get all possible values for a cell."""
|
|
row_values = set(board[row])
|
|
col_values = set(board[i][col] for i in range(9))
|
|
|
|
# Get filled values in the current 3x3 subgrid
|
|
box_row, box_col = 3 * (row // 3), 3 * (col // 3)
|
|
box_values = set()
|
|
for i in range(box_row, box_row + 3):
|
|
for j in range(box_col, box_col + 3):
|
|
box_values.add(board[i][j])
|
|
|
|
used_values = row_values | col_values | box_values
|
|
return set(range(1, 10)) - used_values
|
|
|
|
def _solve(self, board: list[list[int]]) -> bool:
|
|
"""Solve sudoku using backtracking"""
|
|
empty = self._find_empty(board)
|
|
if not empty:
|
|
return True
|
|
|
|
row, col = empty
|
|
for num in self._get_possible_values(board, row, col):
|
|
board[row][col] = num
|
|
if self._solve(board):
|
|
return True
|
|
board[row][col] = 0
|
|
return False
|
|
|
|
def _find_empty(self, board: list[list[int]]) -> Optional[tuple[int, int]]:
|
|
"""Find an empty cell"""
|
|
for i in range(9):
|
|
for j in range(9):
|
|
if board[i][j] == 0:
|
|
return (i, j)
|
|
return None
|
|
|
|
def _generate_solved_board(self, rng: Random) -> list[list[int]]:
|
|
"""Generate a complete solved sudoku board"""
|
|
board = [[0] * 9 for _ in range(9)]
|
|
|
|
# Fill diagonal boxes first (they are independent)
|
|
for i in range(0, 9, 3):
|
|
nums = list(range(1, 10))
|
|
rng.shuffle(nums)
|
|
pos = 0
|
|
for r in range(i, i + 3):
|
|
for c in range(i, i + 3):
|
|
board[r][c] = nums[pos]
|
|
pos += 1
|
|
|
|
# Solve the rest
|
|
self._solve(board)
|
|
return board
|
|
|
|
def _count_solutions(self, board: list[list[int]], limit: int = 2) -> int:
|
|
"""Count the number of solutions for a given board"""
|
|
|
|
def _get_min_possibilities_cell(board: list[list[int]]) -> Optional[tuple[int, int, set[int]]]:
|
|
"""
|
|
Get the cell with the lowest number of possibilities.
|
|
Returns None if the board is already solved.
|
|
"""
|
|
min_possibilities = 10
|
|
min_cell = None
|
|
min_values = None
|
|
|
|
for i in range(9):
|
|
for j in range(9):
|
|
if board[i][j] == 0:
|
|
possible = self._get_possible_values(board, i, j)
|
|
if len(possible) < min_possibilities:
|
|
min_possibilities = len(possible)
|
|
min_cell = (i, j)
|
|
min_values = possible
|
|
if min_possibilities == 1:
|
|
return (*min_cell, min_values)
|
|
|
|
return (*min_cell, min_values) if min_cell else None
|
|
|
|
def _count_solutions_helper(board: list[list[int]]) -> int:
|
|
cell_info = _get_min_possibilities_cell(board)
|
|
if not cell_info:
|
|
return 1
|
|
|
|
row, col, possible_values = cell_info
|
|
count = 0
|
|
for num in possible_values:
|
|
board[row][col] = num
|
|
count += _count_solutions_helper(board)
|
|
if count >= limit:
|
|
return count
|
|
board[row][col] = 0
|
|
return count
|
|
|
|
return _count_solutions_helper(board)
|
|
|
|
def _create_puzzle(self, solved_board: list[list[int]], num_empty: int, rng: Random) -> list[list[int]]:
|
|
"""Create puzzle by removing numbers from solved board"""
|
|
puzzle = [row[:] for row in solved_board]
|
|
cells = [(i, j) for i in range(9) for j in range(9)]
|
|
rng.shuffle(cells)
|
|
num_removed = 0
|
|
|
|
for i, j in cells:
|
|
saved = puzzle[i][j]
|
|
puzzle[i][j] = 0
|
|
puzzle_copy = copy.deepcopy(puzzle)
|
|
# Check if removing this clue breaks uniqueness
|
|
if self._count_solutions(puzzle_copy) > 1:
|
|
puzzle[i][j] = saved
|
|
else:
|
|
num_removed += 1
|
|
if num_removed == num_empty:
|
|
break
|
|
|
|
return puzzle
|
|
|
|
def _board_to_string(self, board: list[list[int]]) -> str:
|
|
"""Convert board to string representation"""
|
|
return "\n".join(" ".join(str(x) if x != 0 else "_" for x in row) for row in board)
|
|
|
|
def __getitem__(self, idx: int) -> dict:
|
|
"""Generate a single sudoku puzzle"""
|
|
rng = Random(self.seed + idx)
|
|
|
|
# Generate solved board
|
|
solved_board = self._generate_solved_board(rng)
|
|
|
|
# Create puzzle by removing numbers
|
|
num_empty = rng.randint(self.config.min_empty, self.config.max_empty)
|
|
puzzle = self._create_puzzle(solved_board, num_empty, rng)
|
|
|
|
# Format as strings
|
|
puzzle_str = self._board_to_string(puzzle)
|
|
solution_str = self._board_to_string(solved_board)
|
|
|
|
question = (
|
|
f"Solve this Sudoku puzzle:\n{puzzle_str}\n"
|
|
"Respond with only your answer, formatted as the puzzle, a 9x9 grid with numbers separated by spaces, and rows separated by newlines."
|
|
)
|
|
|
|
return {
|
|
"question": question,
|
|
"answer": solution_str,
|
|
"metadata": {
|
|
"puzzle": puzzle,
|
|
"solution": solved_board,
|
|
"num_empty": num_empty,
|
|
"difficulty": {
|
|
"empty": (self.config.min_empty, self.config.max_empty),
|
|
},
|
|
},
|
|
}
|
|
|
|
def score_answer(self, answer: Optional[str], entry: dict[str, Any]) -> float:
|
|
if not isinstance(answer, str) or len(answer) == 0:
|
|
return 0.0
|
|
|
|
oracle_answer = entry["answer"]
|
|
metadata = entry["metadata"]
|
|
solution: list[list[int]] = metadata["solution"]
|
|
board_size: int = len(solution[0])
|
|
|
|
# 1. match answer without trailing whitespaces
|
|
answer_stripped = "\n".join(l.rstrip() for l in answer.split("\n"))
|
|
oracle_answer_stripped = "\n".join(l.rstrip() for l in oracle_answer.split("\n"))
|
|
|
|
if answer_stripped == oracle_answer_stripped:
|
|
reward = 1.0
|
|
else:
|
|
# 2. accept answers with correct numeric sequence (ignoring non-numeric characters)
|
|
row = 0
|
|
num_matching = 0
|
|
for ln in answer.split("\n"):
|
|
if row >= len(solution):
|
|
break
|
|
numbers = [int(c) for c in ln if c in "123456789"]
|
|
if len(numbers) != board_size:
|
|
continue # ignore lines without numbers
|
|
for a, b in zip(solution[row], numbers):
|
|
if a == b:
|
|
num_matching += 1
|
|
row += 1
|
|
|
|
reward = num_matching / (board_size * board_size)
|
|
reward *= 0.9 # penalty for not using standard format
|
|
|
|
if len(answer) > len(oracle_answer):
|
|
reward *= len(oracle_answer) / len(answer) # penalty for additional length
|
|
return reward
|
|
|
|
|
|
class SudokuCurriculum(BaseCurriculum):
|
|
def __init__(self):
|
|
super().__init__(SudokuCurriculum.__name__, SudokuConfig)
|
|
|
|
# Define attributes
|
|
self._define_attributes(
|
|
RangeAttributeDefinition(
|
|
name="empty",
|
|
levels=[20, 30, 40, 50],
|
|
description="Number of empty cells in the puzzle",
|
|
lower_field_name="min_empty",
|
|
upper_field_name="max_empty",
|
|
ensure_interval=True,
|
|
)
|
|
)
|
|
|
|
|
|
register_dataset("sudoku", SudokuDataset, SudokuConfig, SudokuCurriculum)
|