code/train/build_grid_puzzle_training_data.py

#!/usr/bin/env python3
"""Build SPAgent GRPO datasets for Sokoban and FifteenPuzzle.

Initial states are sampled from the registered Beyond-MMEnv environments.
Images and transitions use ``spagent.grid_puzzle_protocol`` so the initial
frame, rollout frames, direct-answer replay, and reward logic cannot drift.

Example smoke build:

    /path/to/mmenv-pool/.venv/bin/python \
        train/build_grid_puzzle_training_data.py \
        --count-per-task 10 \
        --output-dir outputs/grpo_converted/grid_puzzles_v1_smoke

The FifteenPuzzle upstream generator shuffles without a parity check. This
builder rejects unsolvable boards and records only verified-solvable states.
"""

from __future__ import annotations

import argparse
import copy
import json
import random
import sys
from concurrent.futures import ProcessPoolExecutor
from pathlib import Path
from typing import Any, Dict, Iterable, List, Sequence


REPO_ROOT = Path(__file__).resolve().parents[1]
THINK_WITH_IMAGES_ROOT = REPO_ROOT.parents[1]
DEFAULT_BEYOND_ROOT = THINK_WITH_IMAGES_ROOT / "Beyond-MMEnv-Scaling"
DEFAULT_OUTPUT_DIR = REPO_ROOT / "outputs" / "grpo_converted" / "grid_puzzles_v1"

if str(REPO_ROOT) not in sys.path:
    sys.path.insert(0, str(REPO_ROOT))

from spagent.grid_puzzle_protocol import (  # noqa: E402
    FIFTEEN_TASK,
    SOKOBAN_TASK,
    SUPPORTED_TASKS,
    build_system_prompt,
    build_user_prompt,
    canonical_fifteen_state,
    canonical_sokoban_state,
    encode_initial_state,
    render_state,
    state_text,
)

_WORKER_ENV: Any = None


def is_fifteen_solvable(board: Sequence[int], *, size: int = 4) -> bool:
    """Return the standard even-width sliding-puzzle parity test."""
    values = [int(value) for value in board]
    nonzero = [value for value in values if value != 0]
    inversions = sum(
        1
        for i, left in enumerate(nonzero)
        for right in nonzero[i + 1:]
        if left > right
    )
    blank_row_from_bottom = size - (values.index(0) // size)
    if size % 2 == 1:
        return inversions % 2 == 0
    return (blank_row_from_bottom % 2 == 0) != (inversions % 2 == 0)


def _import_gym_v(beyond_root: Path) -> Any:
    pool_root = beyond_root / "mmenv-pool"
    if not pool_root.is_dir():
        raise FileNotFoundError(f"Beyond-MMEnv mmenv-pool not found: {pool_root}")
    if str(pool_root) not in sys.path:
        sys.path.insert(0, str(pool_root))
    try:
        import gym_v
        import gym_v.envs  # noqa: F401
    except ImportError as exc:
        raise RuntimeError(
            "Could not import gym_v. Run this builder with the mmenv-pool "
            "virtualenv, for example:\n"
            f"  {pool_root}/.venv/bin/python {Path(__file__).resolve()}"
        ) from exc
    return gym_v


def _extract_sokoban_state(base: Any) -> Dict[str, Any]:
    room_state = base._ta_env.room_state
    room_fixed = base._ta_env.room_fixed
    rows, cols = room_state.shape
    walls = []
    goals = []
    boxes = []
    players = []
    for r in range(rows):
        for c in range(cols):
            fixed = int(room_fixed[r, c])
            current = int(room_state[r, c])
            if fixed == 0:
                walls.append((r, c))
            if fixed == 2:
                goals.append((r, c))
            if current in (3, 4):
                boxes.append((r, c))
            if current in (5, 6):
                players.append((r, c))
    if len(players) != 1:
        raise RuntimeError(f"Expected exactly one Sokoban player, found {players}")
    return canonical_sokoban_state(
        rows=rows,
        cols=cols,
        walls=walls,
        goals=goals,
        boxes=boxes,
        player=players[0],
    )


def _extract_fifteen_state(base: Any) -> Dict[str, Any]:
    board = [
        0 if value is None else int(value)
        for row in base._ta_env.board
        for value in row
    ]
    return canonical_fifteen_state(board)


def sample_state(env: Any, task_id: str, seed: int) -> Dict[str, Any]:
    env.reset(seed=seed)
    base = env.unwrapped
    if task_id == SOKOBAN_TASK:
        return _extract_sokoban_state(base)
    if task_id == FIFTEEN_TASK:
        return _extract_fifteen_state(base)
    raise ValueError(f"Unsupported task_id: {task_id}")


def _init_sampling_worker(beyond_root: str, task_id: str) -> None:
    """Create one reusable environment per sampling worker."""
    global _WORKER_ENV
    gym_v = _import_gym_v(Path(beyond_root))
    try:
        from loguru import logger

        logger.remove()
    except (ImportError, ValueError):
        pass
    _WORKER_ENV = gym_v.make(task_id)


def _sample_seed_batch(
    task_id: str,
    seed_start: int,
    seed_stop: int,
) -> List[tuple[int, Dict[str, Any] | None, str | None]]:
    if _WORKER_ENV is None:
        raise RuntimeError("Sampling worker environment was not initialized")
    results: List[tuple[int, Dict[str, Any] | None, str | None]] = []
    for seed in range(seed_start, seed_stop):
        try:
            state = sample_state(_WORKER_ENV, task_id, seed)
        except Exception as exc:
            results.append(
                (seed, None, f"{type(exc).__name__}: {exc}")
            )
        else:
            results.append((seed, state, None))
    return results


def _task_slug(task_id: str) -> str:
    return "sokoban" if task_id == SOKOBAN_TASK else "fifteen_puzzle"


def build_record(
    *,
    task_id: str,
    state: Dict[str, Any],
    seed: int,
    index: int,
    output_dir: Path,
) -> Dict[str, Any]:
    slug = _task_slug(task_id)
    image_path = output_dir / "images" / slug / f"{slug}_{index:06d}.png"
    render_state(task_id, state, image_path)
    encoded = encode_initial_state(state)
    return {
        "id": f"{slug}_{index:06d}",
        "task_id": task_id,
        "source_env": task_id,
        "source_seed": seed,
        "messages": [
            {"role": "system", "content": build_system_prompt(task_id)},
            {
                "role": "user",
                "content": build_user_prompt(task_id, state, image_path),
            },
        ],
        "images": [str(image_path.resolve())],
        "initial_state": encoded,
        "state_text": state_text(task_id, state),
        "solvable": True,
        "deterministic": True,
        "fully_observed": True,
    }


def collect_records(
    *,
    gym_v: Any,
    task_id: str,
    count: int,
    start_seed: int,
    output_dir: Path,
    max_seed_attempts: int,
    beyond_root: Path,
    workers: int,
    sample_batch_size: int,
) -> List[Dict[str, Any]]:
    records: List[Dict[str, Any]] = []
    seen = set()
    attempts = 0

    def consume(
        seed: int,
        state: Dict[str, Any] | None,
        error: str | None,
    ) -> None:
        nonlocal attempts
        attempts += 1
        if error is not None or state is None:
            print(
                f"[skip] {task_id} seed={seed}: "
                f"{error or 'missing state'}",
                flush=True,
            )
            return
        encoded = encode_initial_state(state)
        if encoded in seen:
            return
        if (
            task_id == FIFTEEN_TASK
            and not is_fifteen_solvable(state["board"])
        ):
            return
        seen.add(encoded)
        records.append(
            build_record(
                task_id=task_id,
                state=state,
                seed=seed,
                index=len(records),
                output_dir=output_dir,
            )
        )
        if len(records) % 100 == 0 or len(records) == count:
            print(
                f"[{task_id}] built {len(records)}/{count} "
                f"(scanned {attempts} seeds)",
                flush=True,
            )

    if workers == 1:
        seed = start_seed
        env = gym_v.make(task_id)
        try:
            while len(records) < count and attempts < max_seed_attempts:
                try:
                    state = sample_state(env, task_id, seed)
                except Exception as exc:
                    consume(
                        seed,
                        None,
                        f"{type(exc).__name__}: {exc}",
                    )
                else:
                    consume(seed, state, None)
                seed += 1
        finally:
            env.close()
    else:
        next_seed = start_seed
        with ProcessPoolExecutor(
            max_workers=workers,
            initializer=_init_sampling_worker,
            initargs=(str(beyond_root), task_id),
        ) as executor:
            while len(records) < count and attempts < max_seed_attempts:
                remaining = max_seed_attempts - attempts
                window = min(
                    remaining,
                    workers * sample_batch_size * 4,
                )
                ranges = []
                seed_stop = next_seed + window
                for batch_start in range(
                    next_seed, seed_stop, sample_batch_size
                ):
                    ranges.append(
                        (
                            task_id,
                            batch_start,
                            min(batch_start + sample_batch_size, seed_stop),
                        )
                    )
                next_seed = seed_stop
                batches = executor.map(
                    _sample_seed_batch,
                    *(zip(*ranges)),
                    chunksize=1,
                )
                done = False
                for batch in batches:
                    for seed, state, error in batch:
                        if (
                            len(records) >= count
                            or attempts >= max_seed_attempts
                        ):
                            done = True
                            break
                        consume(seed, state, error)
                    if done:
                        break

    if len(records) != count:
        raise RuntimeError(
            f"Built only {len(records)}/{count} records for {task_id} after "
            f"{attempts} seed attempts"
        )
    return records


def _write_jsonl(path: Path, records: Iterable[Dict[str, Any]]) -> None:
    path.parent.mkdir(parents=True, exist_ok=True)
    with path.open("w", encoding="utf-8") as handle:
        for record in records:
            handle.write(json.dumps(record, ensure_ascii=False) + "\n")


def split_records(
    records: Sequence[Dict[str, Any]],
    *,
    test_ratio: float,
    split_seed: int,
) -> tuple[List[Dict[str, Any]], List[Dict[str, Any]]]:
    shuffled = list(records)
    random.Random(split_seed).shuffle(shuffled)
    if not shuffled:
        return [], []
    test_count = int(len(shuffled) * test_ratio)
    if test_ratio > 0 and len(shuffled) > 1:
        test_count = max(1, test_count)
    test_count = min(test_count, max(0, len(shuffled) - 1))
    return shuffled[test_count:], shuffled[:test_count]


def pad_records(
    records: Sequence[Dict[str, Any]],
    *,
    count: int,
    seed: int,
) -> List[Dict[str, Any]]:
    """Deterministically repeat records when the source has fewer unique states."""
    padded = list(records)
    if len(padded) >= count:
        return padded[:count]
    if not padded:
        raise ValueError("Cannot pad an empty record collection")
    pool = list(records)
    random.Random(seed).shuffle(pool)
    while len(padded) < count:
        source = pool[(len(padded) - len(records)) % len(pool)]
        repeated = copy.deepcopy(source)
        slug = _task_slug(str(source["task_id"]))
        repeated["id"] = f"{slug}_{len(padded):06d}"
        repeated["duplicate_of"] = source["id"]
        padded.append(repeated)
    return padded


def parse_args() -> argparse.Namespace:
    parser = argparse.ArgumentParser(description=__doc__)
    parser.add_argument(
        "--beyond-root",
        type=Path,
        default=DEFAULT_BEYOND_ROOT,
        help=f"Beyond-MMEnv-Scaling root (default: {DEFAULT_BEYOND_ROOT})",
    )
    parser.add_argument(
        "--output-dir",
        type=Path,
        default=DEFAULT_OUTPUT_DIR,
        help=f"Dataset output directory (default: {DEFAULT_OUTPUT_DIR})",
    )
    parser.add_argument(
        "--task",
        action="append",
        choices=SUPPORTED_TASKS,
        help="Task to build; repeat for multiple tasks. Defaults to both.",
    )
    parser.add_argument("--count-per-task", type=int, default=2000)
    parser.add_argument(
        "--sokoban-count",
        type=int,
        default=None,
        help="Override --count-per-task for Games/Sokoban-v0.",
    )
    parser.add_argument(
        "--fifteen-count",
        type=int,
        default=None,
        help="Override --count-per-task for Games/FifteenPuzzle-v0.",
    )
    parser.add_argument(
        "--sokoban-unique-count",
        type=int,
        default=1600,
        help=(
            "Unique Sokoban states to collect before deterministic padding. "
            "The official fixed 6x6 generator has fewer than 2000 practical "
            "unique states."
        ),
    )
    parser.add_argument("--test-ratio", type=float, default=0.0)
    parser.add_argument("--start-seed", type=int, default=0)
    parser.add_argument("--split-seed", type=int, default=0)
    parser.add_argument("--max-seed-attempts", type=int, default=100000)
    parser.add_argument(
        "--workers",
        type=int,
        default=16,
        help="Parallel sampling workers; use 1 for serial generation.",
    )
    parser.add_argument(
        "--sample-batch-size",
        type=int,
        default=16,
        help="Consecutive seeds sampled per worker job.",
    )
    return parser.parse_args()


def main() -> None:
    args = parse_args()
    if args.count_per_task <= 0:
        raise SystemExit("--count-per-task must be positive")
    if args.sokoban_count is not None and args.sokoban_count <= 0:
        raise SystemExit("--sokoban-count must be positive")
    if args.fifteen_count is not None and args.fifteen_count <= 0:
        raise SystemExit("--fifteen-count must be positive")
    if args.sokoban_unique_count <= 0:
        raise SystemExit("--sokoban-unique-count must be positive")
    if not 0 <= args.test_ratio < 1:
        raise SystemExit("--test-ratio must be in [0, 1)")
    if args.workers <= 0:
        raise SystemExit("--workers must be positive")
    if args.sample_batch_size <= 0:
        raise SystemExit("--sample-batch-size must be positive")
    tasks = args.task or list(SUPPORTED_TASKS)
    output_dir = args.output_dir.resolve()
    output_dir.mkdir(parents=True, exist_ok=True)
    gym_v = _import_gym_v(args.beyond_root.resolve())

    combined_train: List[Dict[str, Any]] = []
    combined_test: List[Dict[str, Any]] = []
    manifest_tasks: Dict[str, Any] = {}
    for task_offset, task_id in enumerate(tasks):
        if task_id == SOKOBAN_TASK:
            task_count = args.sokoban_count or args.count_per_task
            collection_count = min(task_count, args.sokoban_unique_count)
        else:
            task_count = args.fifteen_count or args.count_per_task
            collection_count = task_count
        records = collect_records(
            gym_v=gym_v,
            task_id=task_id,
            count=collection_count,
            start_seed=args.start_seed,
            output_dir=output_dir,
            max_seed_attempts=args.max_seed_attempts,
            beyond_root=args.beyond_root.resolve(),
            workers=args.workers,
            sample_batch_size=args.sample_batch_size,
        )
        unique_initial_states = len(
            {str(record["initial_state"]) for record in records}
        )
        records = pad_records(
            records,
            count=task_count,
            seed=args.split_seed + task_offset,
        )
        train, test = split_records(
            records,
            test_ratio=args.test_ratio,
            split_seed=args.split_seed + task_offset,
        )
        slug = _task_slug(task_id)
        _write_jsonl(output_dir / f"{slug}_all_v1.jsonl", records)
        _write_jsonl(output_dir / f"{slug}_train_v1.jsonl", train)
        _write_jsonl(output_dir / f"{slug}_test_v1.jsonl", test)
        combined_train.extend(train)
        combined_test.extend(test)
        manifest_tasks[task_id] = {
            "records": len(records),
            "train": len(train),
            "test": len(test),
            "unique_initial_states": unique_initial_states,
            "repeated_records": len(records) - unique_initial_states,
            "seed_range_scanned_from": args.start_seed,
        }

    random.Random(args.split_seed).shuffle(combined_train)
    random.Random(args.split_seed + 1).shuffle(combined_test)
    _write_jsonl(output_dir / "grid_puzzles_train_v1.jsonl", combined_train)
    _write_jsonl(output_dir / "grid_puzzles_test_v1.jsonl", combined_test)
    manifest = {
        "format": "spagent-grid-puzzles-v1",
        "tasks": manifest_tasks,
        "combined_train": len(combined_train),
        "combined_test": len(combined_test),
        "test_ratio": args.test_ratio,
        "sampling_workers": args.workers,
        "fully_observed": True,
        "deterministic": True,
        "fifteen_filter": "standard sliding-puzzle parity",
    }
    (output_dir / "manifest.json").write_text(
        json.dumps(manifest, indent=2, ensure_ascii=False) + "\n",
        encoding="utf-8",
    )
    print(f"DONE: {output_dir}")
    print(json.dumps(manifest, indent=2, ensure_ascii=False))


if __name__ == "__main__":
    main()