Datasets:

naivecynics
/

MetaScore-SFT-100

"""Step 2: serialize indexed piano MIDI into sharded 16-bar JSONL.
Every JSONL line contains one non-overlapping 16-counted-bar window:
    {
        "content": [{"start": 16, "pitch": 60, "duration": 4}],
        "split": 80,
        "barlines": [16, 32, ..., 272],
    }
The first four reconstructed bars precede ``split`` and the remaining twelve
bars follow it. ``barlines`` contains the 17 window-local boundaries of those
16 counted bars, while locations 0--15 remain reserved for an optional pickup.
Bar boundaries are reconstructed from MIDI time-signature events using the same
snapping behaviour observed when MuseScore imports MIDI: a time-signature
change is snapped backwards to the latest completed barline.  If the first
sounding note would not start at location 16, the first reconstructed bar is
treated as an uncounted pickup bar.
"""

from __future__ import annotations

import argparse
import bisect
import csv
import gzip
import hashlib
import json
import os
import sys
import tempfile
from collections import defaultdict, deque
from concurrent.futures import ProcessPoolExecutor
from dataclasses import dataclass
from fractions import Fraction
from pathlib import Path
from typing import Iterable, Iterator, Sequence

import mido

from .config import (
    DATASETS,
    INDEX_DIR,
    PIPELINE_VERSION,
    RAW_DIR,
    REQUIRED_CATEGORIES,
    REQUIRE_4_4,
    SLICES_DIR,
    SOURCES_PER_SHARD,
    WORKERS,
)


PIANO_PROGRAMS = frozenset(range(8))
PICKUP_STEPS = 16
WINDOW_BARS = 16
PROMPT_BARS = 4


class SerializationError(ValueError):
    """Raised when a MIDI cannot be serialized without silently repairing it."""


@dataclass(frozen=True, order=True)
class Note:
    start: int
    end: int
    pitch: int


@dataclass(frozen=True)
class TimeSignature:
    tick: int
    numerator: int
    denominator: int


@dataclass(frozen=True)
class ParsedMidi:
    ticks_per_beat: int
    notes: tuple[Note, ...]
    time_signatures: tuple[TimeSignature, ...]


@dataclass(frozen=True)
class PianoPartSelection:
    extraction_mode: str
    selected_port: int | None
    selected_channel: int | None
    selected_note_count: int
    selected_active_bars: int | None


@dataclass(frozen=True)
class FileResult:
    input_path: Path
    output_path: Path
    bar_count: int
    window_count: int
    note_count: int


@dataclass(frozen=True)
class DirectoryResult:
    files: tuple[FileResult, ...]
    failures: tuple[tuple[Path, str], ...]

    @property
    def total_windows(self) -> int:
        return sum(result.window_count for result in self.files)


@dataclass(frozen=True)
class _TimedEvent:
    tick: int
    track_index: int
    event_index: int
    port: int
    message: mido.Message | mido.MetaMessage


@dataclass(frozen=True)
class _Window:
    start_bar: int
    anchor_bar: int
    split_bar: int
    end_bar: int


def _normalize_time_signatures(
    events: Iterable[tuple[int, int, int]]
) -> tuple[TimeSignature, ...]:
    by_tick: dict[int, tuple[int, int]] = {}
    for tick, numerator, denominator in events:
        if numerator <= 0 or denominator <= 0:
            raise SerializationError(
                f"invalid time signature at tick {tick}: "
                f"{numerator}/{denominator}"
            )
        signature = (numerator, denominator)
        previous = by_tick.get(tick)
        if previous is not None and previous != signature:
            raise SerializationError(
                f"conflicting time signatures at tick {tick}: "
                f"{previous[0]}/{previous[1]} and "
                f"{numerator}/{denominator}"
            )
        by_tick[tick] = signature

    if 0 not in by_tick:
        by_tick[0] = (4, 4)

    return tuple(
        TimeSignature(tick, numerator, denominator)
        for tick, (numerator, denominator) in sorted(by_tick.items())
    )


def parse_midi(path: Path) -> ParsedMidi:
    """Strictly parse one single-piano MIDI into paired notes and meter events."""
    try:
        midi = mido.MidiFile(path, clip=False)
    except Exception as exc:
        raise SerializationError(f"cannot parse MIDI: {exc}") from exc

    if midi.type == 2:
        raise SerializationError("MIDI type 2 contains asynchronous tracks")
    if midi.ticks_per_beat <= 0:
        raise SerializationError(
            f"invalid ticks_per_beat: {midi.ticks_per_beat}"
        )

    timed_events: list[_TimedEvent] = []
    signature_events: list[tuple[int, int, int]] = []

    for track_index, track in enumerate(midi.tracks):
        tick = 0
        port = 0
        for event_index, message in enumerate(track):
            tick += message.time
            if tick < 0:
                raise SerializationError(
                    f"negative absolute tick in track {track_index}"
                )
            if message.type == "midi_port":
                port = message.port
            elif message.type == "time_signature":
                signature_events.append(
                    (tick, message.numerator, message.denominator)
                )

            if message.type in {"program_change", "note_on", "note_off"}:
                timed_events.append(
                    _TimedEvent(
                        tick=tick,
                        track_index=track_index,
                        event_index=event_index,
                        port=port,
                        message=message,
                    )
                )

    timed_events.sort(
        key=lambda event: (
            event.tick,
            event.track_index,
            event.event_index,
        )
    )

    programs: defaultdict[tuple[int, int], int] = defaultdict(int)
    active: defaultdict[
        tuple[int, int, int, int], deque[int]
    ] = defaultdict(deque)
    audible_parts: set[tuple[int, int]] = set()
    notes: list[Note] = []

    for event in timed_events:
        message = event.message
        part = (event.port, message.channel)

        if message.type == "program_change":
            programs[part] = message.program
            continue

        key = (
            event.track_index,
            event.port,
            message.channel,
            message.note,
        )
        is_note_on = message.type == "note_on" and message.velocity > 0

        if is_note_on:
            if message.channel == 9:
                raise SerializationError(
                    f"drum note at tick {event.tick} is not allowed"
                )
            program = programs[part]
            if program not in PIANO_PROGRAMS:
                raise SerializationError(
                    f"non-piano program {program} sounds at tick {event.tick}"
                )
            audible_parts.add(part)
            if len(audible_parts) > 1:
                raise SerializationError(
                    "multiple piano logical parts are not allowed: "
                    + ", ".join(
                        f"port={port},channel={channel}"
                        for port, channel in sorted(audible_parts)
                    )
                )
            active[key].append(event.tick)
            continue

        starts = active.get(key)
        if not starts:
            raise SerializationError(
                f"unmatched note-off for pitch {message.note} "
                f"at tick {event.tick} in track {event.track_index}"
            )
        start = starts.popleft()
        if not starts:
            del active[key]
        if event.tick <= start:
            raise SerializationError(
                f"non-positive note duration for pitch {message.note} "
                f"at tick {start}"
            )
        notes.append(Note(start, event.tick, message.note))

    if active:
        key, starts = next(iter(active.items()))
        raise SerializationError(
            f"unmatched note-on for pitch {key[3]} at tick {starts[0]} "
            f"in track {key[0]}"
        )
    if not notes:
        raise SerializationError("MIDI contains no sounding notes")

    return ParsedMidi(
        ticks_per_beat=midi.ticks_per_beat,
        notes=tuple(sorted(notes)),
        time_signatures=_normalize_time_signatures(signature_events),
    )


def parse_dominant_piano_part(
    path: Path,
) -> tuple[ParsedMidi, PianoPartSelection]:
    """Extract the dominant pure-piano logical part from an ensemble MIDI.
    A logical part is a ``(port, channel)`` pair.  Parts that sound under any
    non-piano program are excluded.  Candidates are ranked by sounding note
    count, then active-bar coverage, then ascending port/channel.
    """
    try:
        midi = mido.MidiFile(path, clip=False)
    except Exception as exc:
        raise SerializationError(f"cannot parse MIDI: {exc}") from exc
    if midi.type == 2:
        raise SerializationError("MIDI type 2 contains asynchronous tracks")
    if midi.ticks_per_beat <= 0:
        raise SerializationError(
            f"invalid ticks_per_beat: {midi.ticks_per_beat}"
        )

    timed_events: list[_TimedEvent] = []
    signature_events: list[tuple[int, int, int]] = []
    for track_index, track in enumerate(midi.tracks):
        tick = 0
        port = 0
        for event_index, message in enumerate(track):
            tick += message.time
            if tick < 0:
                raise SerializationError(
                    f"negative absolute tick in track {track_index}"
                )
            if message.type == "midi_port":
                port = message.port
            elif message.type == "time_signature":
                signature_events.append(
                    (tick, message.numerator, message.denominator)
                )
            if message.type in {"program_change", "note_on", "note_off"}:
                timed_events.append(
                    _TimedEvent(
                        tick=tick,
                        track_index=track_index,
                        event_index=event_index,
                        port=port,
                        message=message,
                    )
                )
    timed_events.sort(
        key=lambda event: (
            event.tick,
            event.track_index,
            event.event_index,
        )
    )

    programs: defaultdict[tuple[int, int], int] = defaultdict(int)
    sounding_programs: defaultdict[tuple[int, int], set[int]] = defaultdict(set)
    onset_ticks: defaultdict[tuple[int, int], list[int]] = defaultdict(list)
    for event in timed_events:
        message = event.message
        part = (event.port, message.channel)
        if message.type == "program_change":
            programs[part] = message.program
            continue
        if message.type != "note_on" or message.velocity == 0:
            continue
        if message.channel == 9:
            continue
        program = programs[part]
        sounding_programs[part].add(program)
        onset_ticks[part].append(event.tick)

    candidates = [
        part
        for part, part_programs in sounding_programs.items()
        if part_programs and part_programs.issubset(PIANO_PROGRAMS)
    ]
    if not candidates:
        raise SerializationError("MIDI contains no pure piano logical part")

    signatures = _normalize_time_signatures(signature_events)
    last_onset = max(tick for part in candidates for tick in onset_ticks[part])
    coverage_source = ParsedMidi(
        ticks_per_beat=midi.ticks_per_beat,
        notes=(Note(0, max(1, last_onset + 1), 0),),
        time_signatures=signatures,
    )
    boundaries = reconstruct_bar_boundaries(coverage_source)

    def active_bar_count(part: tuple[int, int]) -> int:
        return len(
            {
                max(0, bisect.bisect_right(boundaries, tick) - 1)
                for tick in onset_ticks[part]
            }
        )

    active_bars = {part: active_bar_count(part) for part in candidates}
    selected = min(
        candidates,
        key=lambda part: (
            -len(onset_ticks[part]),
            -active_bars[part],
            part,
        ),
    )

    active: defaultdict[
        tuple[int, int, int, int], deque[int]
    ] = defaultdict(deque)
    notes: list[Note] = []
    for event in timed_events:
        message = event.message
        part = (event.port, message.channel)
        if message.type == "program_change" or part != selected:
            continue
        key = (
            event.track_index,
            event.port,
            message.channel,
            message.note,
        )
        is_note_on = message.type == "note_on" and message.velocity > 0
        if is_note_on:
            active[key].append(event.tick)
            continue
        starts = active.get(key)
        if not starts:
            raise SerializationError(
                f"unmatched note-off for selected piano pitch "
                f"{message.note} at tick {event.tick}"
            )
        start = starts.popleft()
        if not starts:
            del active[key]
        if event.tick <= start:
            raise SerializationError(
                f"non-positive selected piano note duration for pitch "
                f"{message.note} at tick {start}"
            )
        notes.append(Note(start, event.tick, message.note))

    if active:
        key, starts = next(iter(active.items()))
        raise SerializationError(
            f"unmatched note-on for selected piano pitch {key[3]} "
            f"at tick {starts[0]}"
        )
    if not notes:
        raise SerializationError("selected piano logical part has no notes")
    return (
        ParsedMidi(
            ticks_per_beat=midi.ticks_per_beat,
            notes=tuple(sorted(notes)),
            time_signatures=signatures,
        ),
        PianoPartSelection(
            extraction_mode="dominant_piano_part",
            selected_port=selected[0],
            selected_channel=selected[1],
            selected_note_count=len(onset_ticks[selected]),
            selected_active_bars=active_bars[selected],
        ),
    )


def reconstruct_bar_boundaries(parsed: ParsedMidi) -> tuple[Fraction, ...]:
    """Reconstruct MuseScore-style bar boundaries from time signatures.
    MIDI does not carry explicit barline events.  MuseScore imports a
    time-signature change by snapping it backwards to the most recent complete
    bar under the preceding signature.  Any residual ticks are therefore
    absorbed into the first bar under the new signature rather than counted as
    an extra short measure.
    """
    last_note_end = max(note.end for note in parsed.notes)
    cursor = Fraction(0)
    boundaries: list[Fraction] = [cursor]
    signatures = parsed.time_signatures

    for index, signature in enumerate(signatures):
        stop = (
            Fraction(signatures[index + 1].tick)
            if index + 1 < len(signatures)
            else Fraction(last_note_end)
        )
        if stop < cursor:
            raise SerializationError(
                f"time-signature change at tick {stop} precedes "
                f"the reconstructed bar cursor {cursor}"
            )

        bar_length = Fraction(
            parsed.ticks_per_beat * 4 * signature.numerator,
            signature.denominator,
        )
        if bar_length <= 0:
            raise SerializationError(
                f"invalid bar length for "
                f"{signature.numerator}/{signature.denominator}"
            )

        complete_bars = int((stop - cursor) // bar_length)
        for _ in range(complete_bars):
            cursor += bar_length
            boundaries.append(cursor)

    if cursor < last_note_end:
        final_signature = signatures[-1]
        final_bar_length = Fraction(
            parsed.ticks_per_beat * 4 * final_signature.numerator,
            final_signature.denominator,
        )
        cursor += final_bar_length
        boundaries.append(cursor)

    return tuple(boundaries)


def round_half_up(value: Fraction) -> int:
    """Round halves away from zero using exact rational arithmetic."""
    if value < 0:
        return -round_half_up(-value)
    return (2 * value.numerator + value.denominator) // (
        2 * value.denominator
    )


def ticks_to_steps(ticks: Fraction, ticks_per_beat: int) -> int:
    return round_half_up(ticks * 4 / ticks_per_beat)


def has_pickup(parsed: ParsedMidi) -> bool:
    """Apply the project heuristic for an uncounted opening pickup bar."""
    first_note_tick = min(note.start for note in parsed.notes)
    ordinary_start = PICKUP_STEPS + ticks_to_steps(
        Fraction(first_note_tick),
        parsed.ticks_per_beat,
    )
    return ordinary_start != PICKUP_STEPS


def _window_layout(bar_count: int, pickup: bool) -> tuple[_Window, ...]:
    windows: list[_Window] = []
    next_bar = 0

    if pickup:
        if bar_count >= WINDOW_BARS + 1:
            windows.append(
                _Window(
                    start_bar=0,
                    anchor_bar=1,
                    split_bar=1 + PROMPT_BARS,
                    end_bar=1 + WINDOW_BARS,
                )
            )
        next_bar = WINDOW_BARS + 1

    while next_bar + WINDOW_BARS <= bar_count:
        windows.append(
            _Window(
                start_bar=next_bar,
                anchor_bar=next_bar,
                split_bar=next_bar + PROMPT_BARS,
                end_bar=next_bar + WINDOW_BARS,
            )
        )
        next_bar += WINDOW_BARS

    return tuple(windows)


def serialize_windows(
    parsed: ParsedMidi,
    boundaries: Sequence[Fraction],
    *,
    pickup: bool | None = None,
) -> list[dict[str, object]]:
    """Convert complete non-overlapping 16-bar windows to JSON objects.
    ``pickup`` overrides the MIDI-only heuristic when an authoritative score
    representation (for example MuseScore XML) provides the opening-measure
    length.  Existing training preprocessing keeps the heuristic by omitting
    this argument.
    """
    bar_count = len(boundaries) - 1
    layout = _window_layout(
        bar_count,
        has_pickup(parsed) if pickup is None else pickup,
    )
    notes = parsed.notes
    next_note = 0
    active: list[Note] = []
    windows: list[dict[str, object]] = []

    for window_index, window in enumerate(layout):
        window_start = boundaries[window.start_bar]
        location_anchor = boundaries[window.anchor_bar]
        split_tick = boundaries[window.split_bar]
        window_end = boundaries[window.end_bar]
        window_steps = ticks_to_steps(
            window_end - location_anchor,
            parsed.ticks_per_beat,
        )
        if window_steps <= 0:
            raise SerializationError(
                f"window {window_index} has non-positive quantized length"
            )

        active = [note for note in active if note.end > window_start]
        while (
            next_note < len(notes)
            and notes[next_note].start < window_end
        ):
            note = notes[next_note]
            if note.end > window_start:
                active.append(note)
            next_note += 1

        quantized_notes: set[tuple[int, int, int]] = set()
        for note in active:
            clipped_start = max(Fraction(note.start), window_start)
            clipped_end = min(Fraction(note.end), window_end)
            if clipped_end <= clipped_start:
                continue

            relative_start = ticks_to_steps(
                clipped_start - location_anchor,
                parsed.ticks_per_beat,
            )
            relative_end = ticks_to_steps(
                clipped_end - location_anchor,
                parsed.ticks_per_beat,
            )
            # Nearest-grid quantization can round a very late onset exactly to
            # the right boundary.  Pull it back one step so every serialized
            # note remains inside its 16-bar window.
            relative_start = min(
                max(relative_start, -PICKUP_STEPS),
                window_steps - 1,
            )
            relative_end = min(
                window_steps,
                max(relative_start + 1, relative_end),
            )
            duration = relative_end - relative_start
            quantized_notes.add(
                (
                    PICKUP_STEPS + relative_start,
                    note.pitch,
                    duration,
                )
            )

        content = [
            {"start": start, "pitch": pitch, "duration": duration}
            for start, pitch, duration in sorted(quantized_notes)
        ]
        split = PICKUP_STEPS + ticks_to_steps(
            split_tick - location_anchor,
            parsed.ticks_per_beat,
        )
        barlines = [
            PICKUP_STEPS
            + ticks_to_steps(
                boundaries[bar_index] - location_anchor,
                parsed.ticks_per_beat,
            )
            for bar_index in range(
                window.anchor_bar,
                window.end_bar + 1,
            )
        ]
        if len(barlines) != WINDOW_BARS + 1:
            raise SerializationError(
                f"window {window_index} has {len(barlines)} barlines; "
                f"expected {WINDOW_BARS + 1}"
            )
        if barlines[0] != PICKUP_STEPS:
            raise SerializationError(
                f"window {window_index} starts at location {barlines[0]}; "
                f"expected {PICKUP_STEPS}"
            )
        if any(
            right <= left
            for left, right in zip(barlines, barlines[1:])
        ):
            raise SerializationError(
                f"window {window_index} has non-increasing quantized barlines"
            )
        if barlines[PROMPT_BARS] != split:
            raise SerializationError(
                f"window {window_index} split {split} does not match "
                f"barline {barlines[PROMPT_BARS]}"
            )
        windows.append(
            {
                "content": content,
                "split": split,
                "barlines": barlines,
            }
        )

    return windows


def _write_jsonl_atomic(
    output_path: Path,
    rows: Iterable[dict[str, object]],
) -> None:
    output_path.parent.mkdir(parents=True, exist_ok=True)
    temporary_path: Path | None = None
    try:
        with tempfile.NamedTemporaryFile(
            mode="w",
            encoding="utf-8",
            newline="\n",
            prefix=f".{output_path.name}.",
            suffix=".tmp",
            dir=output_path.parent,
            delete=False,
        ) as output:
            temporary_path = Path(output.name)
            for row in rows:
                json.dump(
                    row,
                    output,
                    ensure_ascii=False,
                    separators=(",", ":"),
                )
                output.write("\n")
            output.flush()
            os.fsync(output.fileno())
        os.replace(temporary_path, output_path)
    except Exception:
        if temporary_path is not None:
            temporary_path.unlink(missing_ok=True)
        raise


def serialize_file(input_path: Path, output_path: Path) -> FileResult:
    parsed = parse_midi(input_path)
    boundaries = reconstruct_bar_boundaries(parsed)
    windows = serialize_windows(parsed, boundaries)
    _write_jsonl_atomic(output_path, windows)
    return FileResult(
        input_path=input_path,
        output_path=output_path,
        bar_count=len(boundaries) - 1,
        window_count=len(windows),
        note_count=len(parsed.notes),
    )


def _midi_paths(input_dir: Path) -> list[Path]:
    paths = sorted(
        path
        for path in input_dir.iterdir()
        if path.is_file() and path.suffix.lower() in {".mid", ".midi"}
    )
    stems: dict[str, Path] = {}
    for path in paths:
        previous = stems.get(path.stem)
        if previous is not None:
            raise SerializationError(
                f"output filename collision: {previous.name} and {path.name}"
            )
        stems[path.stem] = path
    return paths


def serialize_directory(
    input_dir: Path,
    output_dir: Path,
) -> DirectoryResult:
    if not input_dir.is_dir():
        raise FileNotFoundError(f"input directory not found: {input_dir}")
    output_dir.mkdir(parents=True, exist_ok=True)

    successes: list[FileResult] = []
    failures: list[tuple[Path, str]] = []
    for input_path in _midi_paths(input_dir):
        output_path = output_dir / f"{input_path.stem}.jsonl"
        try:
            result = serialize_file(input_path, output_path)
        except Exception as exc:
            failures.append((input_path, str(exc)))
            print(f"[error] {input_path.name}: {exc}", file=sys.stderr)
            continue
        successes.append(result)
        print(
            f"[ok] {input_path.name}: "
            f"{result.bar_count} bars, "
            f"{result.window_count} windows, "
            f"{result.note_count} notes"
        )

    return DirectoryResult(tuple(successes), tuple(failures))


def _eligible(row: dict[str, str]) -> bool:
    return (
        row["parse_status"] == "ok"
        and row["category"] in REQUIRED_CATEGORIES
        and (not REQUIRE_4_4 or row["all_4_4"] == "1")
    )


def iter_index_rows(
    index_dir: Path,
    dataset: str,
    limit: int | None = None,
) -> Iterator[dict[str, str]]:
    if limit == 0:
        return
    path = index_dir / f"{dataset}.csv.gz"
    if not path.is_file():
        raise FileNotFoundError(f"index not found: {path}")

    count = 0
    with gzip.open(path, "rt", encoding="utf-8", newline="") as source:
        for row in csv.DictReader(source):
            if not _eligible(row):
                continue
            yield row
            count += 1
            if limit is not None and count >= limit:
                return


def _chunks(
    rows: Iterable[dict[str, str]],
    size: int,
) -> Iterator[list[dict[str, str]]]:
    chunk: list[dict[str, str]] = []
    for row in rows:
        chunk.append(row)
        if len(chunk) == size:
            yield chunk
            chunk = []
    if chunk:
        yield chunk


def _dataset_root(dataset: str, raw_dir: Path) -> Path:
    if dataset == "all":
        return raw_dir
    if dataset == "notagen":
        return raw_dir / "notagen_pretrain-" / "midi"
    if dataset == "chongchong":
        return raw_dir / "chongchong-" / "midi"
    raise ValueError(f"unknown dataset: {dataset}")


def _serialize_index_row(
    task: tuple[dict[str, str], str],
) -> tuple[list[dict[str, object]], dict[str, object] | None]:
    row, dataset_root_text = task
    input_path = Path(dataset_root_text) / row["relative_path"]
    try:
        if row["category"] == "single_piano":
            parsed = parse_midi(input_path)
            selection = PianoPartSelection(
                extraction_mode="full_single_piano",
                selected_port=None,
                selected_channel=None,
                selected_note_count=len(parsed.notes),
                selected_active_bars=None,
            )
        elif row["category"] == "contains_piano":
            parsed, selection = parse_dominant_piano_part(input_path)
        else:
            raise SerializationError(
                f"unsupported indexed category: {row['category']}"
            )
        boundaries = reconstruct_bar_boundaries(parsed)
        windows = serialize_windows(parsed, boundaries)
    except Exception as exc:
        return [], {
            "sample_id": row["sample_id"],
            "dataset": row["dataset"],
            "relative_path": row["relative_path"],
            "error": str(exc),
            "error_type": type(exc).__name__,
        }

    metadata = {
        "sample_id": row["sample_id"],
        "dataset": row["dataset"],
        "source_partition": row["source_partition"],
        "relative_path": row["relative_path"],
        "sha256": row["sha256"],
        "source_category": row["category"],
        "extraction_mode": selection.extraction_mode,
        "selected_port": selection.selected_port,
        "selected_channel": selection.selected_channel,
        "selected_piano_note_count": selection.selected_note_count,
        "selected_piano_active_bars": selection.selected_active_bars,
    }
    slices = [
        {
            **metadata,
            "slice_id": f"{row['sample_id']}#{window_index:06d}",
            "window_index": window_index,
            **window,
        }
        for window_index, window in enumerate(windows)
    ]
    return slices, None


def _write_json_atomic(path: Path, value: dict[str, object]) -> None:
    path.parent.mkdir(parents=True, exist_ok=True)
    temporary = path.with_name(f".{path.name}.tmp")
    temporary.write_text(
        json.dumps(value, indent=2, sort_keys=True) + "\n",
        encoding="utf-8",
    )
    os.replace(temporary, path)


def _shard_complete(
    output_path: Path,
    errors_path: Path,
    done_path: Path,
    input_sha256: str,
) -> bool:
    if (
        not output_path.is_file()
        or not errors_path.is_file()
        or not done_path.is_file()
    ):
        return False
    try:
        metadata = json.loads(done_path.read_text(encoding="utf-8"))
    except (OSError, json.JSONDecodeError):
        return False
    return (
        metadata.get("pipeline_version") == PIPELINE_VERSION
        and metadata.get("input_sha256") == input_sha256
    )


def _chunk_checksum(rows: Sequence[dict[str, str]]) -> str:
    digest = hashlib.sha256()
    for row in rows:
        digest.update(row["sample_id"].encode("utf-8"))
        digest.update(b"\0")
        digest.update(row["sha256"].encode("ascii"))
        digest.update(b"\0")
    return digest.hexdigest()


def build_slices(
    *,
    raw_dir: Path,
    index_dir: Path,
    output_dir: Path,
    datasets: Sequence[str],
    workers: int,
    sources_per_shard: int,
    limit: int | None = None,
) -> dict[str, int]:
    output_dir.mkdir(parents=True, exist_ok=True)
    summary = {"sources": 0, "slices": 0, "errors": 0, "shards": 0}
    executor = ProcessPoolExecutor(max_workers=workers) if workers > 1 else None

    try:
        for dataset in datasets:
            dataset_root = _dataset_root(dataset, raw_dir)
            rows = iter_index_rows(index_dir, dataset, limit)
            for shard_index, chunk in enumerate(_chunks(rows, sources_per_shard)):
                stem = f"{dataset}-{shard_index:06d}"
                output_path = output_dir / f"{stem}.jsonl"
                errors_path = output_dir / f"{stem}.errors.jsonl"
                done_path = output_dir / f"{stem}.done.json"
                input_sha256 = _chunk_checksum(chunk)
                if _shard_complete(
                    output_path,
                    errors_path,
                    done_path,
                    input_sha256,
                ):
                    metadata = json.loads(done_path.read_text(encoding="utf-8"))
                    for key in summary:
                        summary[key] += int(metadata.get(key, 0))
                    print(f"[resume] {stem}")
                    continue

                tasks = ((row, str(dataset_root)) for row in chunk)
                results = (
                    executor.map(_serialize_index_row, tasks, chunksize=16)
                    if executor is not None
                    else map(_serialize_index_row, tasks)
                )
                slices: list[dict[str, object]] = []
                errors: list[dict[str, object]] = []
                for source_slices, error in results:
                    slices.extend(source_slices)
                    if error is not None:
                        errors.append(error)

                shard_summary = {
                    "sources": len(chunk),
                    "slices": len(slices),
                    "errors": len(errors),
                    "shards": 1,
                    "pipeline_version": PIPELINE_VERSION,
                    "input_sha256": input_sha256,
                }
                _write_jsonl_atomic(output_path, slices)
                _write_jsonl_atomic(errors_path, errors)
                _write_json_atomic(done_path, shard_summary)
                for key in summary:
                    summary[key] += int(shard_summary[key])
                print(
                    f"[ok] {stem}: {len(chunk)} sources, "
                    f"{len(slices)} slices, {len(errors)} errors"
                )
    finally:
        if executor is not None:
            executor.shutdown()

    _write_json_atomic(output_dir / "summary.json", summary)
    return summary


def parse_args(argv: Sequence[str] | None = None) -> argparse.Namespace:
    parser = argparse.ArgumentParser(description=__doc__)
    parser.add_argument(
        "--raw-dir",
        type=Path,
        default=RAW_DIR,
    )
    parser.add_argument(
        "--index-dir",
        type=Path,
        default=INDEX_DIR,
    )
    parser.add_argument("--output-dir", type=Path, default=SLICES_DIR)
    parser.add_argument(
        "--datasets",
        nargs="+",
        choices=("all", "notagen", "chongchong"),
        default=DATASETS,
    )
    parser.add_argument("--workers", type=int, default=WORKERS)
    parser.add_argument(
        "--sources-per-shard",
        type=int,
        default=SOURCES_PER_SHARD,
    )
    parser.add_argument("--limit", type=int)
    return parser.parse_args(argv)


def main(argv: Sequence[str] | None = None) -> int:
    args = parse_args(argv)
    if args.workers < 1:
        raise ValueError("--workers must be at least 1")
    if args.sources_per_shard < 1:
        raise ValueError("--sources-per-shard must be at least 1")
    if args.limit is not None and args.limit < 0:
        raise ValueError("--limit must be non-negative")

    summary = build_slices(
        raw_dir=args.raw_dir.resolve(),
        index_dir=args.index_dir.resolve(),
        output_dir=args.output_dir.resolve(),
        datasets=tuple(dict.fromkeys(args.datasets)),
        workers=args.workers,
        sources_per_shard=args.sources_per_shard,
        limit=args.limit,
    )
    json.dump(summary, sys.stdout, indent=2, sort_keys=True)
    sys.stdout.write("\n")
    return 0


if __name__ == "__main__":
    raise SystemExit(main())