Blame for scripts/prepare.py

#!/usr/bin/env python3
"""Build Hugging Face–ready Parquet files from the raw VK CSV exports.

Input  (raw/<category>_posts.csv): one row per post plus one row per comment;
       comment rows repeat all fields of their parent post.
Output (data/<category>_<posts|comments>/train.parquet): six configs,
       posts and comments split into separate tables linked by (group, post_id).

Processing:
  * posts and comments are separated, exact duplicates by ID are dropped;
  * rows with empty text (media-only posts, sticker comments) are dropped;
  * float-looking IDs/counters from the comment rows are cast to integers;
  * dates are rebuilt from the Unix timestamps in UTC (the *_date columns in the
    raw CSV are in the collector's local time);
  * personal data in text is masked: user mentions, links to user profiles,
    phone numbers, e-mail addresses;
  * a heuristic language label is added (same rules as tuvan_detector.py
    in the tuvan-lang-in-vk-posts project).

Usage:  python scripts/prepare.py            (from the repository root)
"""
from __future__ import annotations

import csv
import json
import re
import sys
from pathlib import Path

import pandas as pd
import pyarrow as pa
import pyarrow.parquet as pq

ROOT = Path(__file__).resolve().parent.parent
RAW_DIR = ROOT / "raw"
DATA_DIR = ROOT / "data"

SOURCES = {
    "official_media": "official_media_posts.csv",
    "gov_institutions": "gov_institutions_posts.csv",
    "community_media": "community_media_posts.csv",
}

csv.field_size_limit(sys.maxsize)

# --------------------------------------------------------------------------- #
# Personal-data masking
# --------------------------------------------------------------------------- #
USER_MENTION = re.compile(r"\[id\d+\|[^\]]*\]")                       # [id123|Имя]
GROUP_MENTION = re.compile(r"\[(?:club|public|event)\d+\|([^\]]*)\]")  # [club1|Минздрав] -> Минздрав
USER_LINK = re.compile(r"(?:https?://)?(?:m\.)?vk\.com/id\d+\b")
EMAIL = re.compile(r"\b[\w.+-]+@[\w-]+(?:\.[\w-]+)+\b")
PHONE = re.compile(
    r"(?<![\d\w])(?:\+7|8)[\s\-(]*\d{3}[\s\-)]*\d{3}[\s\-]*\d{2}[\s\-]*\d{2}(?![\d\w])"
    r"|(?<![\d\w+])9\d{2}[\s\-]?\d{3}[\s\-]?\d{2}[\s\-]?\d{2}(?![\d\w])"
)


def mask(text: str) -> str:
    text = USER_MENTION.sub("@user", text)
    text = GROUP_MENTION.sub(r"\1", text)
    text = USER_LINK.sub("<vk_user>", text)
    text = EMAIL.sub("<email>", text)
    text = PHONE.sub("<phone>", text)
    return text


# --------------------------------------------------------------------------- #
# Heuristic language label (rules copied from tuvan_detector.py)
# --------------------------------------------------------------------------- #
TUVAN_CHARS = set("ңөүҢӨҮ")
TUVAN_RU_LAYOUT = re.compile("|".join([
    r"\bмен\b", r"\bсен\b", r"\bбис\b", r"\bсилер\b",
    r"\bчуве\b", r"\bчок\b", r"\bбар\b", r"\bтур\b",
    r"\bболур\b", r"\bдээш\b", r"\bкылыр\b",
    r"\bбистин\b", r"\bмээн\b", r"\bсилернин\b",
    r"\bачамнын\b", r"\bавамнын\b", r"\bачазы\b", r"\bавазы\b",
    r"\bооренир\b", r"\bооренип\b", r"\bооредилге\b", r"\bоршээ\b",
    r"\bог-буле\b", r"\bтыванын\b", r"\bог-буленин\b",
    r"\bог-булезинге\b", r"\bог-булелиг\b",
    r"\bменээ\b", r"\bортек\b", r"\bортээ\b",
    r"\bкожууннун\b", r"\bторуттунген\b",
    r"\bчуректиг\b", r"\bчуректеривиске\b", r"\bчурек\b", r"\bчуректер\b",
    r"\bмонгеде\b", r"\bсоолгу\b", r"\bулегер\b",
    r"\bбригадазынын\b", r"\bоглувустун\b", r"\bсумузунун\b", r"\bозуп\b",
    r"\w+нын\b", r"\w+зы\b", r"\w+ынын\b", r"\w+устун\b", r"\w+узунун\b",
    r"\w+ээ\b", r"\w+иг\b", r"\w+иске\b", r"\w+илге\b", r"\bоо\w+\b",
]))


def lang_heuristic(text: str) -> str:
    if any(ch in TUVAN_CHARS for ch in text):
        return "tyv"            # Tuvan written with ң/ө/ү
    if TUVAN_RU_LAYOUT.search(text.lower()):
        return "tyv_ru_layout"  # probably Tuvan typed on a Russian keyboard (noisy)
    return "ru"                 # Russian or other


# --------------------------------------------------------------------------- #
# Reading
# --------------------------------------------------------------------------- #
def as_int(value: str) -> int | None:
    value = value.strip()
    return int(float(value)) if value else None


def read_raw(path: Path, category: str) -> tuple[pd.DataFrame, pd.DataFrame, dict]:
    posts, comments = [], []
    with path.open(encoding="utf-8-sig", newline="") as f:
        for row in csv.DictReader(f):
            if row["type"] == "post":
                posts.append({
                    "post_id": as_int(row["post_id"]),
                    "group": row["group"],
                    "category": category,
                    "text": row["post_text"],
                    "likes": as_int(row["post_likes"]),
                    "comments_count": as_int(row["post_comments_count"]),
                    "date_unix": as_int(row["post_date_unix"]),
                })
            elif row["type"] == "comment":
                comments.append({
                    "comment_id": as_int(row["comment_id"]),
                    "post_id": as_int(row["post_id"]),
                    "group": row["group"],
                    "category": category,
                    "text": row["comment_text"],
                    "likes": as_int(row["comment_likes"]),
                    "date_unix": as_int(row["comment_date_unix"]),
                })
            else:
                raise ValueError(f"{path.name}: unknown row type {row['type']!r}")

    stats = {"raw_posts": len(posts), "raw_comments": len(comments)}
    return pd.DataFrame(posts), pd.DataFrame(comments), stats


def finalize(df: pd.DataFrame, id_col: str, stats: dict, kind: str) -> pd.DataFrame:
    n = len(df)
    df = df.drop_duplicates(["group", id_col], keep="last")
    stats[f"{kind}_duplicates_dropped"] = n - len(df)

    n = len(df)
    df = df[df["text"].str.strip() != ""]
    stats[f"{kind}_empty_dropped"] = n - len(df)

    df = df.copy()
    df["text"] = df["text"].map(mask)
    df["lang_heuristic"] = df["text"].map(lang_heuristic)
    df["date"] = pd.to_datetime(df["date_unix"], unit="s", utc=True)
    df["year"] = df["date"].dt.year.astype("int16")
    df = df.drop(columns="date_unix").sort_values(["date", id_col], ascending=False)
    stats[kind] = len(df)
    stats[f"{kind}_lang"] = df["lang_heuristic"].value_counts().to_dict()
    return df.reset_index(drop=True)


POST_SCHEMA = pa.schema([
    ("post_id", pa.int64()),
    ("group", pa.string()),
    ("category", pa.string()),
    ("text", pa.string()),
    ("lang_heuristic", pa.string()),
    ("likes", pa.int64()),
    ("comments_count", pa.int64()),
    ("date", pa.timestamp("s", tz="UTC")),
    ("year", pa.int16()),
])

COMMENT_SCHEMA = pa.schema([
    ("comment_id", pa.int64()),
    ("post_id", pa.int64()),
    ("group", pa.string()),
    ("category", pa.string()),
    ("text", pa.string()),
    ("lang_heuristic", pa.string()),
    ("likes", pa.int64()),
    ("date", pa.timestamp("s", tz="UTC")),
    ("year", pa.int16()),
])


def write(df: pd.DataFrame, schema: pa.Schema, config: str) -> Path:
    out = DATA_DIR / config / "train.parquet"
    out.parent.mkdir(parents=True, exist_ok=True)
    table = pa.Table.from_pandas(df[schema.names], schema=schema, preserve_index=False)
    pq.write_table(table, out, row_group_size=10_000)
    return out


def main() -> None:
    all_stats = {}
    for category, filename in SOURCES.items():
        print(f"→ {filename}")
        posts, comments, stats = read_raw(RAW_DIR / filename, category)
        posts = finalize(posts, "post_id", stats, "posts")
        comments = finalize(comments, "comment_id", stats, "comments")
        for df, schema, kind in ((posts, POST_SCHEMA, "posts"), (comments, COMMENT_SCHEMA, "comments")):
            path = write(df, schema, f"{category}_{kind}")
            print(f"   {path.relative_to(ROOT)}: {len(df):,} rows, {path.stat().st_size / 1e6:.1f} MB")
        all_stats[category] = stats

    (ROOT / "scripts" / "stats.json").write_text(json.dumps(all_stats, indent=2, ensure_ascii=False))
    print(json.dumps(all_stats, indent=2, ensure_ascii=False))


if __name__ == "__main__":
    main()