split_data.py

import copy
import json
from collections import Counter, defaultdict
from pathlib import Path

from sklearn.model_selection import train_test_split


ROOT = Path("/root/knowledgegrapheval/type_prediction_dataset")
ARTIFACTS_DIR = ROOT / "artifacts"
INPUT_PATH = ROOT / "type_predictor_data.jsonl"
TRAIN_PATH = ROOT / "type_predictor_train.jsonl"
VAL_PATH = ROOT / "type_predictor_val.jsonl"
TEST_PATH = ROOT / "type_predictor_test.jsonl"
SUMMARY_PATH = ROOT / "split_summary.json"
ASSIGNMENTS_PATH = ARTIFACTS_DIR / "split_assignments.jsonl"

RANDOM_STATE = 42
TRAIN_RATIO = 0.8
VAL_RATIO = 0.1
TEST_RATIO = 0.1


def load_rows() -> list[dict]:
    with INPUT_PATH.open(encoding="utf-8") as handle:
        return [json.loads(line) for line in handle if line.strip()]


def dump_jsonl(path: Path, rows: list[dict]) -> None:
    with path.open("w", encoding="utf-8") as handle:
        for row in rows:
            handle.write(json.dumps(row, ensure_ascii=False) + "\n")


def extract_entity_from_spans(row: dict) -> tuple[str, str]:
    sentence = row["sentence"]
    entity_from_chars = sentence[row["start_char"] : row["end_char"]]
    entity_from_tokens = " ".join(sentence.split(" ")[row["start_token"] : row["end_token"] + 1])
    return entity_from_chars, entity_from_tokens


def stratified_split(rows: list[dict]) -> tuple[list[dict], list[dict], list[dict]]:
    labels = [row["type"] for row in rows]
    indices = list(range(len(rows)))

    train_idx, holdout_idx = train_test_split(
        indices,
        test_size=(1.0 - TRAIN_RATIO),
        stratify=labels,
        random_state=RANDOM_STATE,
        shuffle=True,
    )

    holdout_labels = [labels[idx] for idx in holdout_idx]
    val_idx, test_idx = train_test_split(
        holdout_idx,
        test_size=0.5,
        stratify=holdout_labels,
        random_state=RANDOM_STATE,
        shuffle=True,
    )

    train_rows = [copy.deepcopy(rows[idx]) for idx in train_idx]
    val_rows = [copy.deepcopy(rows[idx]) for idx in val_idx]
    test_rows = [copy.deepcopy(rows[idx]) for idx in test_idx]
    return train_rows, val_rows, test_rows


def add_eval_categories(train_rows: list[dict], eval_rows: list[dict]) -> list[dict]:
    train_sentences = {row["sentence"] for row in train_rows}
    output = []
    for row in eval_rows:
        new_row = copy.deepcopy(row)
        if new_row["sentence"] in train_sentences:
            new_row["evaluation_category"] = "seen_sentence_new_entity"
        else:
            new_row["evaluation_category"] = "unseen_sentence"
        output.append(new_row)
    return output


def count_types(rows: list[dict]) -> Counter:
    return Counter(row["type"] for row in rows)


def count_type_and_category(rows: list[dict]) -> dict[str, dict[str, int]]:
    counts = defaultdict(lambda: {"unseen_sentence": 0, "seen_sentence_new_entity": 0})
    for row in rows:
        counts[row["type"]][row["evaluation_category"]] += 1
    return dict(sorted(counts.items()))


def write_split_assignments(
    train_rows: list[dict],
    val_rows: list[dict],
    test_rows: list[dict],
) -> None:
    ARTIFACTS_DIR.mkdir(parents=True, exist_ok=True)
    sentence_groups = []
    for split_name, rows in [("train", train_rows), ("validation", val_rows), ("test", test_rows)]:
        by_sentence = defaultdict(list)
        for row in rows:
            by_sentence[row["sentence"]].append(row)
        for sentence, sentence_rows in by_sentence.items():
            type_counts = Counter(row["type"] for row in sentence_rows)
            sentence_groups.append(
                {
                    "sentence": sentence,
                    "assigned_split": split_name,
                    "row_count": len(sentence_rows),
                    "type_counts": dict(sorted(type_counts.items())),
                    "row_ids": [row["id"] for row in sentence_rows],
                }
            )

    sentence_groups.sort(key=lambda item: (item["assigned_split"], item["sentence"]))
    dump_jsonl(ASSIGNMENTS_PATH, sentence_groups)


def validate(
    original_rows: list[dict],
    train_rows: list[dict],
    val_rows: list[dict],
    test_rows: list[dict],
) -> dict:
    errors = []
    all_rows = train_rows + val_rows + test_rows
    original_by_id = {row["id"]: row for row in original_rows}
    seen_ids = set()

    for split_name, rows in [("train", train_rows), ("validation", val_rows), ("test", test_rows)]:
        for row in rows:
            row_id = row["id"]
            if row_id in seen_ids:
                errors.append(f"duplicate row id across splits: {row_id}")
            seen_ids.add(row_id)

            if row_id not in original_by_id:
                errors.append(f"row id missing from original dataset: {row_id}")
                continue

            baseline = original_by_id[row_id]
            compare_keys = sorted(set(row.keys()) | set(baseline.keys()) - {"evaluation_category"})
            for key in compare_keys:
                if key == "evaluation_category":
                    continue
                if row.get(key) != baseline.get(key):
                    errors.append(f"{split_name} row {row_id} changed original field {key}")
                    break

            chars_entity, tokens_entity = extract_entity_from_spans(row)
            if chars_entity != row["entity"]:
                errors.append(f"{split_name} row {row_id} char span mismatch")
            if tokens_entity != row["entity"]:
                errors.append(f"{split_name} row {row_id} token span mismatch")

            if split_name == "train":
                if "evaluation_category" in row:
                    errors.append(f"train row {row_id} should not have evaluation_category")
            else:
                if row.get("evaluation_category") not in {"unseen_sentence", "seen_sentence_new_entity"}:
                    errors.append(f"{split_name} row {row_id} missing valid evaluation_category")

    if len(original_rows) != len(all_rows):
        errors.append("row count mismatch after splitting")
    if len(original_by_id) != len(seen_ids):
        errors.append("not all row ids are present exactly once")

    all_types = sorted({row["type"] for row in original_rows})
    for split_name, rows in [("train", train_rows), ("validation", val_rows), ("test", test_rows)]:
        split_types = {row["type"] for row in rows}
        missing_types = sorted(set(all_types) - split_types)
        if missing_types:
            errors.append(f"{split_name} missing types: {missing_types}")

    repeat_train, repeat_val, repeat_test = stratified_split(original_rows)
    repeat_val = add_eval_categories(repeat_train, repeat_val)
    repeat_test = add_eval_categories(repeat_train, repeat_test)
    if [row["id"] for row in repeat_train] != [row["id"] for row in train_rows]:
        errors.append("train split is not deterministic for the fixed seed")
    if [row["id"] for row in repeat_val] != [row["id"] for row in val_rows]:
        errors.append("validation split is not deterministic for the fixed seed")
    if [row["id"] for row in repeat_test] != [row["id"] for row in test_rows]:
        errors.append("test split is not deterministic for the fixed seed")

    return {
        "ok": not errors,
        "errors": errors,
    }


def make_summary(
    original_rows: list[dict],
    train_rows: list[dict],
    val_rows: list[dict],
    test_rows: list[dict],
    validation_result: dict,
) -> dict:
    original_type_counts = count_types(original_rows)
    train_type_counts = count_types(train_rows)
    val_type_counts = count_types(val_rows)
    test_type_counts = count_types(test_rows)

    def split_block(name: str, rows: list[dict]) -> dict:
        return {
            "name": name,
            "rows": len(rows),
            "percentage": len(rows) / len(original_rows),
            "unique_sentences": len({row["sentence"] for row in rows}),
            "type_counts": dict(sorted(count_types(rows).items())),
        }

    summary = {
        "random_seed": RANDOM_STATE,
        "requested_split_ratios": {
            "train": TRAIN_RATIO,
            "validation": VAL_RATIO,
            "test": TEST_RATIO,
        },
        "totals": {
            "original_rows": len(original_rows),
            "train_rows": len(train_rows),
            "validation_rows": len(val_rows),
            "test_rows": len(test_rows),
        },
        "splits": {
            "train": split_block("train", train_rows),
            "validation": split_block("validation", val_rows),
            "test": split_block("test", test_rows),
        },
        "per_type_counts": {},
        "evaluation_category_counts": {
            "validation": dict(sorted(Counter(row["evaluation_category"] for row in val_rows).items())),
            "test": dict(sorted(Counter(row["evaluation_category"] for row in test_rows).items())),
        },
        "evaluation_category_counts_by_type": {
            "validation": count_type_and_category(val_rows),
            "test": count_type_and_category(test_rows),
        },
        "validation": validation_result,
    }

    for entity_type in sorted(original_type_counts):
        summary["per_type_counts"][entity_type] = {
            "original": original_type_counts[entity_type],
            "train": train_type_counts[entity_type],
            "validation": val_type_counts[entity_type],
            "test": test_type_counts[entity_type],
            "train_ratio": train_type_counts[entity_type] / original_type_counts[entity_type],
            "validation_ratio": val_type_counts[entity_type] / original_type_counts[entity_type],
            "test_ratio": test_type_counts[entity_type] / original_type_counts[entity_type],
        }

    return summary


def main() -> None:
    original_rows = load_rows()
    ARTIFACTS_DIR.mkdir(parents=True, exist_ok=True)
    train_rows, val_rows, test_rows = stratified_split(original_rows)
    val_rows = add_eval_categories(train_rows, val_rows)
    test_rows = add_eval_categories(train_rows, test_rows)

    dump_jsonl(TRAIN_PATH, train_rows)
    dump_jsonl(VAL_PATH, val_rows)
    dump_jsonl(TEST_PATH, test_rows)
    write_split_assignments(train_rows, val_rows, test_rows)

    validation_result = validate(original_rows, train_rows, val_rows, test_rows)
    summary = make_summary(original_rows, train_rows, val_rows, test_rows, validation_result)
    SUMMARY_PATH.write_text(json.dumps(summary, ensure_ascii=False, indent=2), encoding="utf-8")

    print(json.dumps(summary["totals"], ensure_ascii=False, indent=2))
    print(json.dumps(summary["evaluation_category_counts"], ensure_ascii=False, indent=2))
    if not validation_result["ok"]:
        raise SystemExit("split validation failed")


if __name__ == "__main__":
    main()