open-thought/reasoning-gym
1
0
Fork 0
mirror of https://github.com/open-thought/reasoning-gym.git synced 2026-04-23 16:55:05 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions
reasoning-gym/reasoning_gym/games/knight_swap.py
Andreas Köpf 5d7fbac0ad
Minor question template & score_answer improvements (#261) 
* math prompt improvements
* ignore brackets in complex_arithmetic results
* improve additional instruction in prompt of polynomial_equations
* more strict tests for score_answer in polynomial_equations
* simplify special reward handling
* fix test_intermediate_integration
* fix sokoban dataset
* add common dataset score_answer consistency test
2025-03-04 21:55:09 +01:00

395 lines
16 KiB
Python
Raw Blame History

import collections
import json
from dataclasses import dataclass
from random import Random
from typing import FrozenSet, Optional
from ..factory import ProceduralDataset, register_dataset
QUESTION_TEMPLATE = """Knight Swap Challenge:
```
{board}
```
Legend:
- 'w' = White Knight
- 'B' = Black Knight
- Empty squares are marked with '.'
Objective:
Swap the positions of all white knights with all black knights through valid moves.
Rules:
1. Knights move in L-shape (2 squares + 1 square perpendicular)
2. Knights can only move to empty squares
3. {start_turn} moves first, then players alternate
4. All knights must reach their target positions (white ↔ black)
Question:
Is it possible to swap all knights' positions? If yes, list the moves.
Answer Format:
- For impossible puzzles: "No"
- For possible puzzles: List moves as ["color,from,to", ...]
Example: ["w,A1,B3"] means white knight moves A1→B3
"""
@dataclass
class KnightSwapConfig:
"""Configuration for Knight Swap puzzle generation.
A Knight Swap puzzle involves moving white and black knights on a chess-like board
where each move must be a valid knight's move. The goal is to swap the positions
of white and black knights.
"""
min_nodes: int = 6 # Minimum number of squares on the board
max_nodes: int = 9 # Maximum number of squares on the board
min_pieces: int = 2 # Minimum number of pieces per color
max_pieces: int = 2 # Maximum number of pieces per color
min_steps: int = 4 # Minimum solution length
max_steps: int = 20 # Maximum solution length
max_attempts: int = 100 # Maximum attempts for board generation and puzzle creation
seed: Optional[int] = None
size: int = 5 # Virtual dataset size
impossible_ratio: float = 0.2 # Ratio of puzzles that should be impossible
def validate(self):
"""Validate configuration parameters"""
assert self.min_nodes >= 6, "min_nodes must be >= 6"
assert self.max_nodes >= self.min_nodes, "max_nodes must be >= min_nodes"
assert self.min_pieces >= 1, "min_pieces must be >= 1"
assert self.max_pieces >= self.min_pieces, "max_pieces must be >= min_pieces"
assert self.min_steps >= 1, "min_steps must be >= 1"
assert self.max_steps >= self.min_steps, "max_steps must be >= min_steps"
assert self.max_attempts >= 1, "max_attempts must be >= 1"
assert 0 <= self.impossible_ratio <= 1, "impossible_ratio must be between 0 and 1"
class KnightSwapLogic:
"""Core game logic for Knight Swap puzzles."""
@staticmethod
def is_knight_move(a: str, b: str) -> bool:
"""Check if moving from square 'a' to square 'b' is a legal knight move."""
a_col = ord(a[0].upper()) - ord("A") + 1
a_row = int(a[1:])
b_col = ord(b[0].upper()) - ord("A") + 1
b_row = int(b[1:])
return {abs(a_col - b_col), abs(a_row - b_row)} == {1, 2}
@staticmethod
def is_connected(graph: dict[str, list[str]]) -> bool:
"""Check if a graph is connected (all nodes reachable from any starting node)."""
if not graph:
return True
start = next(iter(graph))
visited = set()
queue = collections.deque([start])
while queue:
node = queue.popleft()
if node not in visited:
visited.add(node)
for neighbor in graph[node]:
if neighbor not in visited:
queue.append(neighbor)
return len(visited) == len(graph)
@staticmethod
def generate_board(num_nodes: int, rng: Random, max_attempts: int = 1000) -> dict[str, list[str]]:
"""Generate a random connected board where edges represent valid knight moves."""
candidates = ["A1", "A2", "A3", "B1", "B2", "B3", "C1", "C2", "C3", "D1", "D2", "D3"]
attempts = 0
while True:
attempts += 1
nodes = rng.sample(candidates, num_nodes)
graph = {node: [] for node in nodes}
for i in range(len(nodes)):
for j in range(i + 1, len(nodes)):
if KnightSwapLogic.is_knight_move(nodes[i], nodes[j]):
graph[nodes[i]].append(nodes[j])
graph[nodes[j]].append(nodes[i])
for node in graph:
graph[node].sort()
if KnightSwapLogic.is_connected(graph):
return graph
if attempts > max_attempts:
raise Exception(f"Failed to generate connected board after {max_attempts} attempts")
@staticmethod
def solve_swap(
board: dict[str, list[str]], pieces: dict[str, str], start_turn: str = "w"
) -> Optional[list[tuple[str, str, str]]]:
"""Find a sequence of moves to swap white and black pieces positions."""
@dataclass(frozen=True)
class GameState:
white_set: FrozenSet[str]
black_set: FrozenSet[str]
turn: str
initial_white = frozenset(pos for pos, piece in pieces.items() if piece == "w")
initial_black = frozenset(pos for pos, piece in pieces.items() if piece == "B")
initial_state = GameState(initial_white, initial_black, start_turn)
queue = collections.deque([initial_state])
visited = {initial_state}
predecessors = {initial_state: (None, None)}
while queue:
state = queue.popleft()
if state.white_set == initial_black and state.black_set == initial_white:
moves = []
cur_state = state
while predecessors[cur_state][0] is not None:
prev_state, move = predecessors[cur_state]
moves.append(move)
cur_state = prev_state
moves.reverse()
return moves
current_positions = state.white_set if state.turn == "w" else state.black_set
for pos in current_positions:
for neighbor in board[pos]:
if neighbor in state.white_set or neighbor in state.black_set:
continue
if state.turn == "w":
new_white = frozenset(p if p != pos else neighbor for p in state.white_set)
new_black = state.black_set
else:
new_black = frozenset(p if p != pos else neighbor for p in state.black_set)
new_white = state.white_set
next_turn = "B" if state.turn == "w" else "w"
new_state = GameState(new_white, new_black, next_turn)
if new_state not in visited:
visited.add(new_state)
predecessors[new_state] = (state, (state.turn, pos, neighbor))
queue.append(new_state)
return None
class KnightSwapDataset(ProceduralDataset):
"""Generates Knight Swap puzzles with configurable parameters."""
def __init__(self, config: KnightSwapConfig):
super().__init__(config=config, seed=config.seed, size=config.size)
self.game_logic = KnightSwapLogic()
def _format_board(self, board: dict[str, list[str]], pieces: dict[str, str]) -> str:
"""Format the board state as a string."""
positions = list(board.keys())
if not positions:
return ""
columns = sorted(set(pos[0] for pos in positions))
rows = sorted(set(int(pos[1:]) for pos in positions), reverse=True)
lines = []
# Header
lines.append(" " + " ".join(columns))
lines.append(" " + "----" * len(columns))
# Board rows
for row in rows:
line = f"{row} |"
for col in columns:
pos = col + str(row)
if pos in pieces:
piece = pieces[pos] if pieces[pos] is not None else "."
line += f" {piece} |"
else:
line += " |"
lines.append(line)
lines.append(" " + "----" * len(columns))
return "\n".join(lines)
def _format_moves(self, moves: list[tuple[str, str, str]]) -> str:
"""Format the solution moves as a string."""
if not moves:
return "No"
return json.dumps([f"{color},{start},{end}" for color, start, end in moves])
def __getitem__(self, idx: int) -> dict:
"""Generate a single Knight Swap puzzle."""
rng = Random(self.seed + idx)
# Keep trying with new boards until we succeed
board_attempts = 0
while board_attempts < self.config.max_attempts:
try:
# Generate a new board
num_nodes = rng.randint(self.config.min_nodes, self.config.max_nodes)
board = self.game_logic.generate_board(num_nodes, rng, max_attempts=self.config.max_attempts)
positions = list(board.keys())
# Decide if this should be an impossible puzzle
make_impossible = rng.random() < self.config.impossible_ratio
# Try different piece placements on this board
for _ in range(50): # Reduced attempts per board since we try multiple boards
# Use fixed number of pieces for more reliable generation
num_pieces = self.config.min_pieces
white_positions = rng.sample(positions, num_pieces)
remaining = [p for p in positions if p not in white_positions]
black_positions = rng.sample(remaining, num_pieces)
pieces = {pos: None for pos in positions}
for pos in white_positions:
pieces[pos] = "w"
for pos in black_positions:
pieces[pos] = "B"
# For impossible puzzles, try a simpler approach: just remove some key connections
board_copy = {k: list(v) for k, v in board.items()} # Make a copy of the board
if make_impossible:
# Remove critical edges that would make the puzzle impossible
critical_edges = []
for w_pos in white_positions:
for b_pos in black_positions:
if b_pos in board_copy[w_pos]:
critical_edges.append((w_pos, b_pos))
if critical_edges: # Only proceed if we found critical edges
# Remove a random critical edge
w_pos, b_pos = rng.choice(critical_edges)
board_copy[w_pos].remove(b_pos)
board_copy[b_pos].remove(w_pos)
# Try both starting turns
for start_turn in ["w", "B"]:
solution = self.game_logic.solve_swap(board_copy, pieces, start_turn)
# Accept solutions with more flexible length requirements
if (make_impossible and solution is None) or (
not make_impossible
and solution is not None
and self.config.min_steps <= len(solution) <= self.config.max_steps
):
board_str = self._format_board(board_copy, pieces)
solution_str = self._format_moves(solution) if solution else "No"
# Generate board states for solvable puzzles
board_states = []
if solution is not None:
current_pieces = dict(pieces)
board_states.append(dict(current_pieces)) # Initial state
for color, start, end in solution:
current_pieces[end] = current_pieces[start]
current_pieces[start] = None
board_states.append(dict(current_pieces))
return {
"question": QUESTION_TEMPLATE.format(board=board_str, start_turn=start_turn),
"answer": solution_str,
"metadata": {
"board": board_copy,
"pieces": pieces,
"start_turn": start_turn,
"solution": solution,
"is_possible": solution is not None,
"num_steps": len(solution) if solution else 0,
"board_states": board_states if solution is not None else None,
},
}
except Exception:
pass # If board generation fails, we'll try again with a new board
board_attempts += 1
raise ValueError(f"Failed to generate valid puzzle after trying {self.config.max_attempts} different boards")
def score_answer(self, answer: Optional[str], entry: dict) -> float:
"""Score the user's solution for the Knight Swap puzzle.
The answer should be either:
- "No" if the puzzle is impossible
- A JSON list of moves in format ["color,start,end", ...] where color is 'w' or 'B'
Returns:
- 1.0 for correct answer (either "No" for impossible puzzles or valid solution of optimal length)
- A proportional score for correct but longer solutions
- 0.05 for valid moves that don't solve the puzzle
- 0.0 for invalid format or None
"""
if not isinstance(answer, str):
return 0.0
answer = answer.strip()
if len(answer) == 0:
return 0.0
# Handle impossible puzzles
if not entry["metadata"]["is_possible"]:
return 1.0 if answer.lower() == "no" else 0.0
# Handle "No" answer for possible puzzles
if answer.lower() == "no":
return 0.0
try:
# Parse moves from JSON list
move_list = json.loads(answer)
if not isinstance(move_list, list):
return 0.0
# Parse moves
moves = []
for move_str in move_list:
color, start, end = move_str.split(",")
if color not in ("w", "B"):
return 0.0
moves.append((color, start, end))
# Validate and apply moves
board = entry["metadata"]["board"]
pieces = dict(entry["metadata"]["pieces"])
current_turn = entry["metadata"]["start_turn"]
# Track board states after each move
board_states = []
board_states.append(dict(pieces)) # Initial state
for color, start, end in moves:
if color != current_turn:
return 0.0
if start not in pieces or pieces[start] != color:
return 0.0
if end not in board[start]:
return 0.0
if end in pieces and pieces[end] is not None:
return 0.0
# Apply move
pieces[end] = pieces[start]
pieces[start] = None
current_turn = "B" if current_turn == "w" else "w"
# Store board state after this move
board_states.append(dict(pieces))
# Check if solved
white_positions = {pos for pos, piece in pieces.items() if piece == "w"}
black_positions = {pos for pos, piece in pieces.items() if piece == "B"}
initial_white = {pos for pos, piece in entry["metadata"]["pieces"].items() if piece == "w"}
initial_black = {pos for pos, piece in entry["metadata"]["pieces"].items() if piece == "B"}
if white_positions == initial_black and black_positions == initial_white:
optimal_moves = len(entry["metadata"]["solution"])
# Add board states to metadata if solution is valid
entry["metadata"]["board_states"] = board_states
if len(moves) <= optimal_moves:
return 1.0
else:
return optimal_moves / len(moves)
return 0.05
except Exception:
return 0.0
register_dataset("knight_swap", KnightSwapDataset, KnightSwapConfig)
Reference in a new issue View git blame Copy permalink