scripts/build_selfplay_corpus.py

"""Build a tiered lc0 self-play position corpus as a VPD1 database.

For every run slice named in lc0_runs.TIERS this streams a bounded prefix of
several tars, converts them to FENs through the rescorer, samples a few
positions per game, and inserts them into the same VPD1 schema the CCRL corpus
uses, so the two are directly comparable.

Two choices are worth stating.

Tars are read as bounded prefixes rather than in full. Positions cost the same
number of bytes either way, but a prefix of each of forty tars spans forty
points in a run's history, where four whole tars span four. Network diversity
per byte downloaded is much better, and it is the diversity that this corpus is
for.

Only a few positions per game are kept. A game contributes about 110 positions,
and consecutive ones are near-duplicates; sampling across phases keeps the
effective sample size close to the row count. The CCRL corpus took three per
game. Eight is the default here because these tars must be downloaded rather
than read from a local archive, and eight cuts the download roughly fourfold
for positions that are still tens of plies apart.

Nothing is dropped for being hard to train on. Chess960 positions are dropped
by default only because their Shredder-FEN castling fields are not standard
chess and the rest of this pipeline assumes standard chess; pass
--keep-chess960 to retain them.
"""

from __future__ import annotations

import argparse
import hashlib
import json
import random
import re
import shutil
import subprocess
import sys
import tempfile
import time
from pathlib import Path

import chess

sys.path.insert(0, str(Path(__file__).resolve().parent.parent / "vex-position-dataset"))
from vpd import (  # noqa: E402
    PHASE_NAMES,
    connect,
    current_counts,
    initialize,
    make_quotas,
    phase_of,
    position_record,
    set_metadata,
)

from lc0_runs import EXCLUDED, TIERS, quotas, run_quotas  # noqa: E402

BASE = "https://storage.lczero.org/files/training_data"
LISTING_ROW = re.compile(r'href="([^"]+\.tar)">[^<]*</a>\s+(\S+)\s+(\S+)\s+(\d+)')
STANDARD_CASTLING = set("KQkq-")
MIN_REAL_TAR = 50_000_000

INSERT_SQL = """
    INSERT OR IGNORE INTO positions(
      random_key, fen, source_split, source_member, game_number, ply,
      result, side_to_move, phase, piece_count, non_pawn_material,
      material_balance, legal_moves, in_check, castling_mask, halfmove_clock
    ) VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
"""


def spread(count: int) -> list[int]:
    """Indices 0..count-1 ordered so that any prefix is spread over the range.

    Van der Corput: reversing the bits of successive integers visits the range
    at ever finer resolution, so stopping early still leaves the tars taken
    scattered across the run's history rather than clustered at one end.
    """

    if count <= 0:
        return []
    bits = max(1, (count - 1).bit_length())
    seen: list[int] = []
    used = set()
    for i in range(1 << bits):
        reversed_bits = int(format(i, f"0{bits}b")[::-1], 2)
        if reversed_bits < count and reversed_bits not in used:
            used.add(reversed_bits)
            seen.append(reversed_bits)
    return seen


def list_tars(run: str, cache: dict[str, list[tuple[str, int]]]) -> list[tuple[str, int]]:
    if run in cache:
        return cache[run]
    out = subprocess.run(
        ["curl", "-sS", "--max-time", "180", f"{BASE}/{run}/"],
        capture_output=True,
        text=True,
    )
    rows = [
        (name, int(size))
        for name, _, _, size in LISTING_ROW.findall(out.stdout)
        if int(size) >= MIN_REAL_TAR
    ]
    cache[run] = rows
    return rows


def fetch_prefix(run: str, tar: str, limit: int, destination: Path) -> int:
    """Download at most `limit` bytes of a tar.

    head closes the socket once it has enough, which the CDN honours where it
    ignores Range requests.
    """

    command = (
        f'curl -sS --max-time 900 "{BASE}/{run}/{tar}" | head -c {limit} > "{destination}"'
    )
    subprocess.run(command, shell=True, capture_output=True)
    return destination.stat().st_size if destination.exists() else 0


def to_plain(rescorer: Path, tar_path: Path, workdir: Path, threads: int) -> Path | None:
    subprocess.run(["tar", "xf", str(tar_path), "-C", str(workdir)], capture_output=True)
    chunks = list(workdir.rglob("*.gz"))
    if not chunks:
        return None
    plain = workdir / "positions.plain"
    subprocess.run(
        [
            str(rescorer),
            "rescore",
            f"--input={chunks[0].parent}",
            "--no-delete-files",
            f"--nnue-plain-file={plain}",
            "--nnue-best-score=true",
            "--nnue-best-move=true",
            "--deblunder=false",
            f"--threads={threads}",
        ],
        capture_output=True,
        text=True,
    )
    return plain if plain.exists() else None


def games(plain: Path):
    """Yield (fen, ply, result) lists, one per game.

    The rescorer appends games back to back and restarts ply at zero for each,
    so a non-increasing ply is a game boundary.
    """

    current: list[tuple[str, int, str]] = []
    fen = None
    ply = None
    last_ply = None
    with open(plain, encoding="utf-8", errors="replace") as handle:
        for row in handle:
            if row.startswith("fen "):
                fen = row[4:].strip()
            elif row.startswith("ply "):
                ply = int(row[4:])
            elif row.startswith("result "):
                result = row[7:].strip()
            elif row.startswith("e") and row.strip() == "e":
                if fen is None or ply is None:
                    continue
                if last_ply is not None and ply <= last_ply and current:
                    yield current
                    current = []
                current.append((fen, ply, result))
                last_ply = ply
                fen = ply = None
    if current:
        yield current


def absolute_result(plain_result: str, white_to_move: bool) -> str:
    """Convert a .plain result to the PGN string VPD1 stores.

    Stockfish's plain format reports the game outcome from the side to move's
    perspective as 1/0/-1. VPD1 stores an absolute PGN result, as the CCRL
    corpus does, so the two are comparable and vpd.py analyze can read them.
    """

    try:
        value = int(plain_result)
    except (TypeError, ValueError):
        return "*"
    if value == 0:
        return "1/2-1/2"
    white_won = (value > 0) == white_to_move
    return "1-0" if white_won else "0-1"


def sample_game(
    positions: list[tuple[str, int, str]],
    per_game: int,
    counts: dict[int, int],
    phase_quotas: dict[int, int],
    rng: random.Random,
    keep_chess960: bool,
) -> list[tuple[chess.Board, int, int, str]]:
    """Pick a spread of positions from one game, respecting phase quotas."""

    by_phase: dict[int, list[tuple[chess.Board, int, str]]] = {p: [] for p in PHASE_NAMES}
    for fen, ply, result in positions:
        if not keep_chess960 and not set(fen.split(" ")[2]) <= STANDARD_CASTLING:
            continue
        try:
            board = chess.Board(fen)
        except ValueError:
            continue
        by_phase[phase_of(board, ply)].append((board, ply, result))

    # Split the per-game budget in the same 15/60/25 proportion as the corpus
    # quota. Taking an equal count from each phase instead saturates the small
    # opening quota long before the others, after which most of a downloaded
    # game is discarded.
    total_quota = sum(phase_quotas.values()) or 1
    chosen: list[tuple[chess.Board, int, int, str]] = []
    for phase, available in by_phase.items():
        if not available or counts[phase] >= phase_quotas[phase]:
            continue
        wanted = max(1, round(per_game * phase_quotas[phase] / total_quota))
        take = min(wanted, len(available), phase_quotas[phase] - counts[phase])
        for board, ply, result in rng.sample(available, take):
            chosen.append((board, phase, ply, result))
    return chosen


def process_tar(
    db,
    rescorer: Path,
    run: str,
    tar: str,
    tier: str,
    limit: int,
    threads: int,
    per_game: int,
    counts: dict[int, int],
    phase_quotas: dict[int, int],
    seed: int,
    keep_chess960: bool,
    scratch: Path,
) -> dict[str, int]:
    stats = {"bytes": 0, "games": 0, "inserted": 0, "chess960": 0}
    workdir = Path(tempfile.mkdtemp(prefix="lc0tar.", dir=scratch))
    try:
        tar_path = workdir / "prefix.tar"
        stats["bytes"] = fetch_prefix(run, tar, limit, tar_path)
        if stats["bytes"] < 100_000:
            return stats
        plain = to_plain(rescorer, tar_path, workdir, threads)
        if plain is None:
            return stats
        tar_path.unlink(missing_ok=True)

        for game_number, positions in enumerate(games(plain), start=1):
            stats["games"] += 1
            stats["chess960"] += sum(
                not set(f.split(" ")[2]) <= STANDARD_CASTLING for f, _, _ in positions
            )
            rng = random.Random(
                int.from_bytes(
                    hashlib.blake2b(
                        f"{seed}:{tar}:{game_number}".encode(), digest_size=8
                    ).digest(),
                    "big",
                )
            )
            for board, phase, ply, result in sample_game(
                positions, per_game, counts, phase_quotas, rng, keep_chess960
            ):
                cursor = db.execute(
                    INSERT_SQL,
                    position_record(
                        board, tier, f"{run}/{tar}", game_number, ply,
                        absolute_result(result, board.turn == chess.WHITE), phase,
                    ),
                )
                if cursor.rowcount:
                    counts[phase] += 1
                    stats["inserted"] += 1
            if all(counts[p] >= phase_quotas[p] for p in phase_quotas):
                break
    finally:
        shutil.rmtree(workdir, ignore_errors=True)
    return stats


def main() -> None:
    parser = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
    parser.add_argument("--output", required=True, type=Path)
    parser.add_argument("--target", type=int, default=2_000_000)
    parser.add_argument("--per-game", type=int, default=12)
    parser.add_argument("--slice-bytes", type=int, default=120_000_000)
    parser.add_argument("--threads", type=int, default=8)
    parser.add_argument("--seed", type=int, default=91)
    parser.add_argument("--keep-chess960", action="store_true")
    parser.add_argument("--scratch", type=Path, default=Path(tempfile.gettempdir()))
    parser.add_argument(
        "--rescorer",
        type=Path,
        default=Path(__file__).resolve().parent.parent / "lc0-rescorer/build/release/rescorer",
    )
    parser.add_argument("--dry-run", action="store_true", help="show the plan and exit")
    args = parser.parse_args()

    if not args.rescorer.exists():
        sys.exit(f"No rescorer at {args.rescorer}")

    tier_quota = quotas(args.target)
    plan: list[tuple[str, str, float, float, int]] = []
    for tier in TIERS:
        for key, quota in run_quotas(tier, tier_quota[tier.name]).items():
            name, window = key.split(":")
            first, last = (float(x) for x in window.split("-"))
            plan.append((tier.name, name, first, last, quota))

    print("Excluded runs:")
    for name, why in EXCLUDED.items():
        print(f"  {name:9} {why}")
    print(f"\nTarget {args.target:,} positions, {args.per_game} per game, "
          f"{args.slice_bytes / 1e6:.0f} MB per tar\n")
    for tier, run, first, last, quota in plan:
        print(f"  {tier:7} {run:9} [{first:.2f}-{last:.2f}]  {quota:>9,}")
    if args.dry_run:
        return

    args.output.parent.mkdir(parents=True, exist_ok=True)
    db = connect(args.output)
    initialize(db)
    set_metadata(db, "corpus", "lc0-selfplay-tiered")
    set_metadata(db, "source", BASE)
    set_metadata(db, "target_positions", args.target)
    set_metadata(db, "seed", args.seed)
    set_metadata(db, "per_game", args.per_game)
    set_metadata(db, "slice_bytes", args.slice_bytes)
    set_metadata(db, "keep_chess960", int(args.keep_chess960))
    set_metadata(db, "excluded_runs", json.dumps(EXCLUDED, sort_keys=True))
    set_metadata(db, "tier_quotas", json.dumps(tier_quota, sort_keys=True))
    db.commit()

    done_row = db.execute("SELECT value FROM metadata WHERE key='tars_done'").fetchone()
    tars_done = set(json.loads(done_row[0])) if done_row else set()

    listing_cache: dict[str, list[tuple[str, int]]] = {}
    started = time.monotonic()

    for tier, run, first, last, quota in plan:
        have = db.execute(
            "SELECT COUNT(*) FROM positions WHERE source_split=? AND source_member LIKE ?",
            (tier, f"{run}/%"),
        ).fetchone()[0]
        if have >= quota:
            print(f"[{tier}/{run}] already at {have:,}/{quota:,}")
            continue

        tars = list_tars(run, listing_cache)
        window = tars[int(len(tars) * first) : max(int(len(tars) * last), 1)]
        if not window:
            print(f"[{tier}/{run}] no tars in window", file=sys.stderr)
            continue

        # Phase quotas are scoped to this run slice so each contributes the same
        # opening/middlegame/endgame mix as the CCRL corpus.
        base = {
            p: db.execute(
                "SELECT COUNT(*) FROM positions WHERE source_split=? "
                "AND source_member LIKE ? AND phase=?",
                (tier, f"{run}/%", p),
            ).fetchone()[0]
            for p in PHASE_NAMES
        }
        slice_quota = make_quotas(quota)
        counts = dict(base)

        print(f"\n[{tier}/{run}] {have:,}/{quota:,} from {len(window)} tars "
              f"in [{first:.2f}-{last:.2f}]")

        for index in spread(len(window)):
            if all(counts[p] >= slice_quota[p] for p in slice_quota):
                break
            tar = window[index][0]
            token = f"{run}/{tar}"
            if token in tars_done:
                continue
            stats = process_tar(
                db, args.rescorer, run, tar, tier, args.slice_bytes, args.threads,
                args.per_game, counts, slice_quota, args.seed, args.keep_chess960,
                args.scratch,
            )
            tars_done.add(token)
            set_metadata(db, "tars_done", json.dumps(sorted(tars_done)))
            db.commit()
            total = sum(counts.values())
            elapsed = time.monotonic() - started
            print(
                f"  {tar[-21:]:21} {stats['bytes'] / 1e6:6.0f}MB "
                f"games={stats['games']:5,} +{stats['inserted']:6,} "
                f"-> {total:,}/{quota:,}  [{elapsed / 60:.1f}m]",
                flush=True,
            )

    total = db.execute("SELECT COUNT(*) FROM positions").fetchone()[0]
    set_metadata(db, "position_count", total)
    set_metadata(db, "completed_unix", int(time.time()))
    db.commit()
    print(f"\nWrote {total:,} positions to {args.output}")
    for row in db.execute(
        "SELECT source_split, COUNT(*) FROM positions GROUP BY source_split ORDER BY 2 DESC"
    ):
        print(f"  {row[0]:8} {row[1]:>9,}")
    db.close()


if __name__ == "__main__":
    main()