scripts/dataset_variants.py

"""Deterministic construction of Sphragis chunk-size configurations."""

from __future__ import annotations

import hashlib
import json
import re
from collections import Counter, defaultdict
from typing import Any

import pyarrow as pa

try:
    from scripts.metrical_lines import load_public_metrical_lines
    from scripts.text_units import encode_text_units, source_text_units
except ModuleNotFoundError:  # Direct execution from the scripts directory.
    from metrical_lines import load_public_metrical_lines
    from text_units import encode_text_units, source_text_units


BASE_CONFIGS = ("prose", "verse_sentence", "verse_metre")
SPLITS = ("train", "validation", "test")
CHUNK_TARGETS = (10, 100)
BOTTLENECK_TARGET = max(CHUNK_TARGETS)
CHUNKING_SEED = 776

CHUNK_FIELDS = (
    pa.field("chunk_size", pa.int64()),
    pa.field("chunk_target_size", pa.int64()),
    pa.field("constituent_ids", pa.list_(pa.string())),
    pa.field("constituent_provenance", pa.string()),
    pa.field("chunk_work_ids", pa.list_(pa.string())),
    pa.field("chunk_works", pa.list_(pa.string())),
    pa.field("chunk_is_mixed_work", pa.bool_()),
)


def variant_schema(base_schema: pa.Schema) -> pa.Schema:
    return pa.schema([*base_schema, *CHUNK_FIELDS], metadata=base_schema.metadata)


def _natural_key(value: Any) -> tuple:
    text = "" if value is None else str(value)
    return tuple(
        (0, int(part)) if part.isdigit() else (1, part.casefold())
        for part in re.split(r"(\d+)", text)
        if part
    )


def actual_order_key(row: dict) -> tuple:
    """Sort a row in its best available work-internal textual order."""
    source_sentence_id = ""
    try:
        records = json.loads(row.get("source_records") or "[]")
        if records:
            source_sentence_id = records[0].get("source_sentence_id", "")
    except (json.JSONDecodeError, TypeError):
        pass
    return (
        _natural_key(row.get("book")),
        _natural_key(row.get("poem_sequence")),
        _natural_key(row.get("line_number")),
        _natural_key(row.get("passage")),
        _natural_key(source_sentence_id),
        row["id"],
    )


def _stable_score(*parts: Any) -> str:
    payload = "\x1f".join(str(part) for part in parts)
    return hashlib.sha256(payload.encode("utf-8")).hexdigest()


def seeded_discard_ids(
    base_config: str, rows: list[dict], target: int, split: str,
) -> list[str]:
    remainder = len(rows) % target
    ranked = sorted(
        rows,
        key=lambda row: _stable_score(
            CHUNKING_SEED, base_config, target, split, row["author"], row["id"],
        ),
    )
    return sorted(row["id"] for row in ranked[:remainder])


def eligible_authors(rows: list[dict], threshold: int) -> set[str]:
    counts = {
        split: Counter(row["author"] for row in rows if row["split"] == split)
        for split in ("validation", "test")
    }
    authors = set(counts["validation"]) | set(counts["test"])
    return {
        author for author in authors
        if counts["validation"][author] >= threshold
        and counts["test"][author] >= threshold
    }


def select_bottleneck_rows(
    base_config: str,
    rows: list[dict],
) -> tuple[list[dict], set[str], dict[str, list[str]]]:
    """Select the atomic corpus shared by every task size for one genre."""
    retained_authors = eligible_authors(rows, BOTTLENECK_TARGET)
    discarded_by_split: dict[str, list[str]] = {"validation": [], "test": []}
    discarded_ids = {
        row["id"] for row in rows if row["author"] not in retained_authors
    }

    for split in ("validation", "test"):
        by_author = defaultdict(list)
        for row in rows:
            if row["split"] == split and row["author"] in retained_authors:
                by_author[row["author"]].append(row)
        for author in sorted(by_author):
            author_discarded = seeded_discard_ids(
                base_config, by_author[author], BOTTLENECK_TARGET, split,
            )
            discarded_by_split[split].extend(author_discarded)
            discarded_ids.update(author_discarded)

    selected = [row for row in rows if row["id"] not in discarded_ids]
    return selected, retained_authors, {
        split: sorted(ids) for split, ids in discarded_by_split.items()
    }


def _ordered_unique(values: list[Any]) -> list[Any]:
    seen = set()
    output = []
    for value in values:
        if value not in seen:
            seen.add(value)
            output.append(value)
    return output


def _provenance(row: dict) -> dict:
    keys = (
        "id", "work", "work_id", "cts_urn", "passage", "treebank_source",
        "book", "poem_sequence", "line_number", "hypotactic_file",
    )
    return {key: row.get(key) for key in keys if key in row}


def _chunk_metadata(row: dict, target: int) -> dict:
    row = dict(row)
    row.update({
        "text": encode_text_units(source_text_units(row["text"])),
        "chunk_size": 1,
        "chunk_target_size": target,
        "constituent_ids": [row["id"]],
        "constituent_provenance": json.dumps(
            [_provenance(row)], ensure_ascii=False, sort_keys=True,
        ),
        "chunk_work_ids": [row["work_id"]],
        "chunk_works": [row["work"]],
        "chunk_is_mixed_work": False,
    })
    return row


def _merge_json_records(rows: list[dict], field: str) -> str:
    merged = []
    seen = set()
    for row in rows:
        for record in json.loads(row[field]):
            key = json.dumps(record, ensure_ascii=False, sort_keys=True)
            if key not in seen:
                seen.add(key)
                merged.append(record)
    return json.dumps(merged, ensure_ascii=False, sort_keys=True)


def _merge_metrical_lines(rows: list[dict]) -> str:
    """Deduplicate overlapping lines by separate IDs, never serialized IDs."""
    merged = []
    seen_ids = set()
    for row in rows:
        lines = load_public_metrical_lines(row["metrical_lines"])
        line_ids = row["metrical_line_ids"]
        if len(lines) != len(line_ids):
            raise ValueError(
                f"metrical line/id count mismatch in row {row['id']}: "
                f"{len(lines)} != {len(line_ids)}"
            )
        for line_id, line in zip(line_ids, lines):
            if line_id not in seen_ids:
                seen_ids.add(line_id)
                merged.append(line)
    return json.dumps(merged, ensure_ascii=False)


def _aggregate_chunk(base_config: str, rows: list[dict], target: int, split: str) -> dict:
    assert len(rows) == target
    authors = {row["author"] for row in rows}
    assert len(authors) == 1
    work_ids = _ordered_unique([row["work_id"] for row in rows])
    works = _ordered_unique([row["work"] for row in rows])
    mixed_work = len(work_ids) > 1
    constituent_ids = [row["id"] for row in rows]
    digest = _stable_score(base_config, target, split, *constituent_ids)[:20]

    chunk = dict(rows[0])
    chunk.update({
        "id": f"chunk-{base_config}-{target}-{digest}",
        "work": works[0] if not mixed_work else "Multiple works",
        "work_id": work_ids[0] if not mixed_work else f"multiple:{digest}",
        "text": encode_text_units([
            unit for row in rows for unit in source_text_units(row["text"])
        ]),
        "conllu": "\n\n".join(row["conllu"].strip() for row in rows) + "\n\n",
        "cts_urn": rows[0]["cts_urn"] if len({row["cts_urn"] for row in rows}) == 1 else None,
        "passage": (
            rows[0]["passage"]
            if len({row["passage"] for row in rows}) == 1
            else f"{rows[0]['passage']}–{rows[-1]['passage']}" if not mixed_work else None
        ),
        "treebank_source": (
            rows[0]["treebank_source"]
            if len({row["treebank_source"] for row in rows}) == 1 else "multiple"
        ),
        "source_records": _merge_json_records(rows, "source_records"),
        "licenses": sorted({license_name for row in rows for license_name in row["licenses"]}),
        "dedup_key": hashlib.sha256("\x1f".join(constituent_ids).encode("utf-8")).hexdigest(),
        "split": split,
        "chunk_size": target,
        "chunk_target_size": target,
        "constituent_ids": constituent_ids,
        "constituent_provenance": json.dumps(
            [_provenance(row) for row in rows], ensure_ascii=False, sort_keys=True,
        ),
        "chunk_work_ids": work_ids,
        "chunk_works": works,
        "chunk_is_mixed_work": mixed_work,
    })

    if base_config == "verse_sentence":
        chunk.update({
            "alignment_component_id": None,
            "component_sentence_index": None,
            "metre": [metre for row in rows for metre in row["metre"]],
            "metrical_line_ids": _ordered_unique([
                line_id for row in rows for line_id in row["metrical_line_ids"]
            ]),
            "metrical_lines": _merge_metrical_lines(rows),
        })
    elif base_config == "verse_metre":
        chunk.update({
            "parent_sentence_ids": _ordered_unique([
                sentence_id for row in rows for sentence_id in row["parent_sentence_ids"]
            ]),
            "alignment_component_id": None,
            "component_line_index": None,
            "book": rows[0]["book"] if len({row["book"] for row in rows}) == 1 else None,
            "poem_sequence": None,
            "line_number": (
                f"{rows[0]['line_number']}–{rows[-1]['line_number']}"
                if not mixed_work else None
            ),
            "metre": "\n".join(row["metre"] for row in rows),
            "syllables": json.dumps(
                [
                    syllable
                    for row in rows
                    for syllable in json.loads(row["syllables"])
                ],
                ensure_ascii=False,
            ),
            "hypotactic_file": (
                rows[0]["hypotactic_file"]
                if len({row["hypotactic_file"] for row in rows}) == 1 else None
            ),
        })
    return chunk


def chunk_author_rows(
    base_config: str,
    rows: list[dict],
    target: int,
    split: str,
) -> tuple[list[dict], list[str]]:
    """Discard the seeded remainder and maximize single-work chunks."""
    assert split in {"validation", "test"}
    discarded_ids = seeded_discard_ids(base_config, rows, target, split)
    discarded = set(discarded_ids)

    by_work = defaultdict(list)
    for row in rows:
        if row["id"] not in discarded:
            by_work[row["work_id"]].append(row)

    chunks = []
    tails = []
    for work_key in sorted(by_work, key=_natural_key):
        ordered = sorted(by_work[work_key], key=actual_order_key)
        full_length = len(ordered) - (len(ordered) % target)
        for start in range(0, full_length, target):
            chunks.append(_aggregate_chunk(base_config, ordered[start:start + target], target, split))
        tails.extend(ordered[full_length:])

    assert len(tails) % target == 0
    for start in range(0, len(tails), target):
        chunks.append(_aggregate_chunk(base_config, tails[start:start + target], target, split))
    return chunks, discarded_ids


def make_dataset_variants(
    rows_by_base_config: dict[str, list[dict]],
) -> tuple[dict[str, list[dict]], dict]:
    variants = {}
    report = {}
    for base_config, rows in rows_by_base_config.items():
        shared_rows, retained_authors, bottleneck_discarded = select_bottleneck_rows(
            base_config, rows,
        )
        variants[f"{base_config}_1"] = [
            {
                **row,
                "text": encode_text_units(source_text_units(row["text"])),
            }
            for row in shared_rows
        ]
        report[f"{base_config}_1"] = {
            "authors": len(retained_authors),
            "retained_authors": sorted(retained_authors),
            "rows": dict(Counter(row["split"] for row in shared_rows)),
            "row_unit": "line" if base_config == "verse_metre" else "sentence",
            "shared_source_selection_target": BOTTLENECK_TARGET,
            "discarded_source_row_ids": bottleneck_discarded,
        }
        for target in CHUNK_TARGETS:
            config = f"{base_config}_{target}"
            variant_rows = [
                _chunk_metadata(row, target)
                for row in shared_rows
                if row["split"] == "train"
            ]
            for split in ("validation", "test"):
                split_chunks = []
                by_author = defaultdict(list)
                for row in shared_rows:
                    if row["split"] == split:
                        by_author[row["author"]].append(row)
                for author in sorted(by_author):
                    chunks, author_discarded = chunk_author_rows(
                        base_config, by_author[author], target, split,
                    )
                    assert not author_discarded
                    split_chunks.extend(chunks)
                variant_rows.extend(split_chunks)

            variants[config] = variant_rows
            split_rows = Counter(row["split"] for row in variant_rows)
            split_source_rows = Counter()
            mixed_work_chunks = Counter()
            for row in variant_rows:
                split_source_rows[row["split"]] += row["chunk_size"]
                if row["chunk_is_mixed_work"]:
                    mixed_work_chunks[row["split"]] += 1
            report[config] = {
                "authors": len(retained_authors),
                "minimum_validation_and_test_source_rows_per_author": BOTTLENECK_TARGET,
                "retained_authors": sorted(retained_authors),
                "shared_source_selection_target": BOTTLENECK_TARGET,
                "train_row_unit": "line" if base_config == "verse_metre" else "sentence",
                "evaluation_row_unit": f"{target}-" + (
                    "line chunk" if base_config == "verse_metre" else "sentence chunk"
                ),
                "rows": dict(split_rows),
                "represented_source_rows": dict(split_source_rows),
                "discarded_source_row_ids": bottleneck_discarded,
                "discarded_source_rows": {
                    split: len(ids) for split, ids in bottleneck_discarded.items()
                },
                "mixed_work_chunks": dict(mixed_work_chunks),
                "chunking_seed": CHUNKING_SEED,
            }
    return variants, report