Upload scripts/limpa_parquet_v76.py with huggingface_hub

scripts/limpa_parquet_v76.py

@@ -0,0 +1,361 @@
+#!/usr/bin/env python3
+import polars as pl
+import re
+import argparse
+import os
+import logging
+import hashlib
+from blingfire import text_to_sentences
+from lingua import Language, LanguageDetectorBuilder
+import nltk
+
+
+# ===================================================================
+# ========================== CONFIG =================================
+# ===================================================================
+
+EXPORT_SPLIT = False
+EXPORT_CSV = False
+
+ADD_ID = True
+ADD_CATEGORY = True
+ADD_SHA = True
+REMOVE_DUPLICATES = True
+
+ROW_LIMIT = 100_000_000
+TEXT_COLUMN = "text"
+
+LAST_ID_FILE = "last_id.txt"       # contém o último ID processado
+FILE_PATTERN = "train-"
+CLEAN_SUFFIX = "_clean_full.parquet"
+
+
+# ===================================================================
+# ====================== NLTK RESOURCES ===============================
+# ===================================================================
+
+def ensure_nltk():
+    required = ["punkt", "punkt_tab"]
+    for r in required:
+        try:
+            nltk.data.find(f"tokenizers/{r}")
+        except LookupError:
+            nltk.download(r, quiet=True)
+ensure_nltk()
+
+
+# ===================================================================
+# ========================= CLASSIFICAÇÃO =============================
+# ===================================================================
+
+INSTRUCT_PREFIX = re.compile(
+    r"^(escreva|explique|resuma|liste|crie|gere|monte|defina|descreva|faça|produza|formule|por favor.*(explique|resuma|liste))",
+    re.IGNORECASE
+)
+QUESTION_PATTERN = re.compile(
+    r"\?$|^(quem|o que|qual|quando|por que|como)\b",
+    re.IGNORECASE
+)
+
+def is_factual(text: str):
+    if not isinstance(text, str):
+        return False
+    return (
+        len(text.split()) > 40
+        and not INSTRUCT_PREFIX.search(text)
+        and not QUESTION_PATTERN.search(text)
+    )
+
+def classify_text(text: str):
+    if not isinstance(text, str):
+        return "other"
+    t = text.strip()
+    if INSTRUCT_PREFIX.search(t):
+        return "instruct"
+    if QUESTION_PATTERN.search(t):
+        return "question"
+    if is_factual(t):
+        return "factual"
+    return "other"
+
+
+# ===================================================================
+# ========================== IDIOMA ==================================
+# ===================================================================
+
+DETECTOR = (
+    LanguageDetectorBuilder
+    .from_languages(Language.PORTUGUESE, Language.ENGLISH, Language.SPANISH)
+    .with_preloaded_language_models()
+    .build()
+)
+
+def is_portuguese(text: str) -> bool:
+    if not isinstance(text, str) or len(text) < 20:
+        return False
+    try:
+        lang = DETECTOR.detect_language_of(text)
+        conf = DETECTOR.compute_language_confidence(text, lang)
+        return lang == Language.PORTUGUESE and conf >= 0.80
+    except:
+        return False
+
+
+# ===================================================================
+# ========================= REGEX / LIMPEZA ===========================
+# ===================================================================
+
+PORN = re.compile(r"pelad|sexo|porn|novinha|xvideos|boquete|anal ", re.IGNORECASE)
+CODE = re.compile(r"function\s*\(|</?script|var\s+|document\.write", re.IGNORECASE)
+SPAM = re.compile(r"bet365|promo|cadastre-se|inscreva|orçamento|ligue ", re.IGNORECASE)
+META = re.compile(r"visualizado|publicado em|conteúdo recente|mostra idioma", re.IGNORECASE)
+
+PATTERNS = [PORN, CODE, SPAM, META]
+
+CPF  = re.compile(r"\b\d{3}\.?\d{3}\.?\d{3}-?\d{2}\b")
+RG   = re.compile(r"\b\d{2}\.?\d{3}\.?\d{3}-?[A-Za-z0-9]{1,2}\b")
+PHONE = re.compile(r"\b(\(?\d{2}\)?\s?)?\d{4,5}-?\d{4}\b")
+EMAIL = re.compile(r"\b[A-Za-z0-9._%+-]+@[A-Za-z0-9.-]+\.[A-Za-z]{2,}\b")
+URL   = re.compile(r"https?://\S+|www\.\S+|\b\S+\.(com|net|org|gov|edu)\b")
+
+HEADINGS = re.compile(
+    r"^(Categoria:|Veja também:|Leia mais:|Últimas notícias:|Para saber mais:|Este artigo|Esta categoria)",
+    re.IGNORECASE
+)
+
+def remove_personal(text):
+    text = CPF.sub("", text)
+    text = RG.sub("", text)
+    text = PHONE.sub("", text)
+    text = EMAIL.sub("", text)
+    text = URL.sub("", text)
+    return text
+
+def remove_patterns_func(text):
+    if not isinstance(text, str):
+        return None
+    for p in PATTERNS:
+        if p.search(text):
+            return None
+    return text
+
+def remove_headings_func(text):
+    return None if HEADINGS.search(text) else text
+
+def normalize(text):
+    if not isinstance(text, str):
+        return None
+    text = re.sub(r'^[\s\u00A0\u1680\u180E\u2000-\u200A\u202F\u205F\u3000]+', '', text)
+    text = remove_personal(text)
+    if remove_headings_func(text) is None:
+        return None
+    text = re.sub(r'\s+', ' ', text)
+    text = re.sub(r'[^\w\s.,!?()\-–]', '', text)
+    text = text.strip()
+    return text if len(text) >= 25 else None
+
+def remove_truncated(text):
+    if not isinstance(text, str):
+        return None
+    if re.match(r"pr[oó]s:|contras:|leia mais|clique|assine", text.lower()):
+        return None
+    return text
+
+def split_sentences(text):
+    raw = text_to_sentences(text)
+    return [s.strip() for s in raw.split("\n") if len(s.strip()) > 25]
+
+def compute_sha(text: str):
+    if not isinstance(text, str):
+        return None
+    return hashlib.sha256(text.encode("utf-8")).hexdigest()
+
+
+# ===================================================================
+# ====================== ID MANAGEMENT ================================
+# ===================================================================
+
+def load_last_id():
+    if not os.path.exists(LAST_ID_FILE):
+        with open(LAST_ID_FILE, "w") as f:
+            f.write("0")
+        return 0
+
+    with open(LAST_ID_FILE, "r") as f:
+        try:
+            return int(f.read().strip())
+        except:
+            return 0
+
+def save_last_id(value):
+    with open(LAST_ID_FILE, "w") as f:
+        f.write(str(value))
+
+
+# ===================================================================
+# ======================= PROCESSAR UM ARQUIVO ========================
+# ===================================================================
+
+def process_file(filepath, id_start):
+
+    filename = os.path.basename(filepath)
+    base = os.path.splitext(filename)[0]
+
+    # =============== Logger individual por arquivo ===================
+    log_name = f"{base}_clean.log"
+
+    logger = logging.getLogger(base)
+    logger.setLevel(logging.INFO)
+
+    fmt = logging.Formatter("%(asctime)s [%(levelname)s] %(message)s")
+
+    fh = logging.FileHandler(log_name, encoding="utf-8")
+    fh.setFormatter(fmt)
+    fh.setLevel(logging.INFO)
+
+    sh = logging.StreamHandler()
+    sh.setFormatter(fmt)
+    sh.setLevel(logging.INFO)
+
+    if logger.hasHandlers():
+        logger.handlers.clear()
+
+    logger.addHandler(fh)
+    logger.addHandler(sh)
+
+    # ================================================================
+    logger.info("\n====================================================")
+    logger.info(f"📦 Processando arquivo: {filename}")
+    logger.info(f"📍 ID inicial: {id_start}")
+    logger.info("====================================================")
+
+    df = pl.read_parquet(filepath)
+
+    if ROW_LIMIT > 0:
+        df = df.head(ROW_LIMIT)
+
+    if TEXT_COLUMN not in df.columns:
+        logger.error(f"❌ Arquivo {filename} sem coluna 'text'. Ignorando.")
+        return id_start
+
+    # ======================= CRIAR IDS ==============================
+    if ADD_ID:
+        new_ids = list(range(id_start, id_start + df.height))
+        df = df.with_columns(pl.Series("id", new_ids))
+        cols = ["id"] + [c for c in df.columns if c != "id"]
+        df = df.select(cols)
+
+        logger.info(f"🆔 ID intervalo: {new_ids[0]} → {new_ids[-1]}")
+
+    removals = {
+        "remove_patterns": [],
+        "remove_lang": [],
+        "remove_normalize": [],
+        "remove_truncated": [],
+        "remove_duplicates_sha": []
+    }
+
+    def track(stage, old, new):
+        old_ids = set(old["id"])
+        new_ids = set(new["id"])
+        removed = sorted(list(old_ids - new_ids))
+        removals[stage].extend(removed)
+        logger.info(f"🗑️ {stage}: {len(removed)} removidos → {removed}")
+
+    # ==================== 1) PATTERNS ==============================
+    old = df
+    df = df.with_columns(pl.col(TEXT_COLUMN).map_elements(remove_patterns_func))
+    df = df.drop_nulls(subset=[TEXT_COLUMN])
+    track("remove_patterns", old, df)
+
+    # ==================== 2) Língua ================================
+    old = df
+    df = df.with_columns(pl.col(TEXT_COLUMN).map_elements(lambda x: x if is_portuguese(x) else None))
+    df = df.drop_nulls(subset=[TEXT_COLUMN])
+    track("remove_lang", old, df)
+
+    # ==================== 3) Normalize =============================
+    old = df
+    df = df.with_columns(pl.col(TEXT_COLUMN).map_elements(normalize))
+    df = df.drop_nulls(subset=[TEXT_COLUMN])
+    track("remove_normalize", old, df)
+
+    # ==================== 4) Truncated =============================
+    old = df
+    df = df.with_columns(pl.col(TEXT_COLUMN).map_elements(remove_truncated))
+    df = df.drop_nulls(subset=[TEXT_COLUMN])
+    track("remove_truncated", old, df)
+
+    # ==================== 5) SHA ==================================
+    if ADD_SHA:
+        df = df.with_columns(pl.col(TEXT_COLUMN).map_elements(compute_sha).alias("sha"))
+
+    # ==================== 6) Duplicados SHA ========================
+    if ADD_SHA and REMOVE_DUPLICATES:
+        old = df
+        df = df.unique(subset=["sha"])
+        track("remove_duplicates_sha", old, df)
+
+    # ==================== 7) CATEGORIA =============================
+    if ADD_CATEGORY:
+        df = df.with_columns(
+            pl.col(TEXT_COLUMN).map_elements(classify_text).alias("category")
+        )
+
+    df = df.sort("id")
+
+    # ==================== Salvar FULL ==============================
+    output_name = f"{base}{CLEAN_SUFFIX}"
+    df.write_parquet(output_name)
+
+    logger.info(f"💾 Salvo: {output_name}")
+
+    # ==================== Atualizar last_id ========================
+    final_last_id = df["id"].max()
+    logger.info(f"🔢 Último ID deste arquivo: {final_last_id}")
+
+    return final_last_id + 1
+
+
+# ===================================================================
+# =============================== MAIN ===============================
+# ===================================================================
+
+def main():
+
+    last_id = load_last_id()
+    next_id = last_id + 1
+
+    # listar arquivos
+    files = [
+        f for f in os.listdir(".")
+        if f.startswith(FILE_PATTERN) and f.endswith(".parquet")
+    ]
+
+    if not files:
+        print("❌ Nenhum parquet encontrado.")
+        return
+
+    # ordenar numericamente
+    def extract_num(fname):
+        try:
+            return int(fname.split("-")[1])
+        except:
+            return 0
+
+    files = sorted(files, key=extract_num)
+
+    print(f"📚 Arquivos encontrados: {files}")
+    print(f"🔢 last_id carregado: {last_id}")
+
+    for file in files:
+        final_id_next = process_file(file, next_id)
+        next_id = final_id_next
+        save_last_id(final_id_next - 1)
+
+    print("\n==================== FINALIZADO (v76) ====================")
+    print(f"🔢 Novo last_id salvo: {next_id - 1}")
+
+
+if __name__ == "__main__":
+    main()