clean.py

#!/usr/bin/env python3
"""
Streaming cleaner for man-page JSONL dataset.
Fixed: robust overstrike removal (bold/underline) that doesn't eat characters.
"""

import json
import re
import sys

# ----------------------------------------------------------------------
# 1. Overstrike removal – SAFE version
# ----------------------------------------------------------------------
def debold(text: str) -> str:
    """
    Remove man-page overstrike sequences only when they follow the classic patterns:
        X\bX  (bold)   →  X
        _\bX  (underline) →  X
    All other backspace sequences are left untouched (shouldn't exist anyway).
    """
    # Pattern 1: (character) \x08 (same character)  -> keep the character
    # Pattern 2: _ \x08 (any character)             -> keep the character
    # The lookahead (?=.) ensures we don't match a trailing backspace with nothing after it.
    text = re.sub(r'(.)\x08(?=\1)', r'\1', text)
    text = re.sub(r'_\x08(.)', r'\1', text)
    # As a final safety net, remove any remaining isolated backspaces
    text = text.replace('\x08', '')
    return text

# ----------------------------------------------------------------------
# 2. Other cleaning helpers (unchanged logic)
# ----------------------------------------------------------------------
def strip_header_footer(text: str) -> str:
    lines = text.splitlines()
    if lines and re.match(r'^[A-Za-z0-9._-]+\([^)]+\)\s+', lines[0]):
        lines.pop(0)
    while lines and lines[-1].strip() == '':
        lines.pop()
    while lines and (
        re.match(r'^[A-Za-z0-9._-]+\([^)]+\)\s*$', lines[-1].strip()) or
        re.match(r'^[A-Z]+\s+\d{4}\s*$', lines[-1].strip())
    ):
        lines.pop()
    return '\n'.join(lines)

def normalize_quotes_dashes(text: str) -> str:
    text = text.replace('``', '“').replace("''", '”')
    return text

def unwrap_paragraphs(text: str) -> str:
    paragraphs = text.split('\n\n')
    new_paras = []
    for para in paragraphs:
        if not para.strip():
            continue
        lines = para.split('\n')
        if all((not line) or line.startswith((' ', '\t')) for line in lines):
            new_paras.append('\n'.join(lines))
        else:
            merged = ' '.join(line.strip() for line in lines if line.strip())
            new_paras.append(merged)
    return '\n\n'.join(new_paras)

def clean_manual(raw_text: str) -> str:
    text = debold(raw_text)                # ← FIXED overstrike removal
    text = strip_header_footer(text)
    text = normalize_quotes_dashes(text)
    text = re.sub(r'\n{3,}', '\n\n', text)
    text = unwrap_paragraphs(text)
    text = re.sub(r'^NAME\n\s+', 'NAME\n', text, flags=re.MULTILINE)
    return text.strip()

# ----------------------------------------------------------------------
# 3. Streaming processor with progress
# ----------------------------------------------------------------------
def process_dataset_streaming(input_path: str, output_cleaned: str, output_training: str = None):
    cleaned_count = 0
    pair_count = 0

    with open(input_path, 'r', encoding='utf-8') as fin, \
         open(output_cleaned, 'w', encoding='utf-8') as fout_cleaned:

        fout_train = None
        if output_training:
            fout_train = open(output_training, 'w', encoding='utf-8')

        try:
            for line_no, line in enumerate(fin, 1):
                line = line.strip()
                if not line:
                    continue
                try:
                    obj = json.loads(line)
                except json.JSONDecodeError as e:
                    print(f"Warning: skipping invalid JSON at line {line_no}: {e}", file=sys.stderr)
                    continue

                topic = obj.get('topic', '')
                section = obj.get('section', '')
                raw = obj.get('manual', '')
                if not raw:
                    continue

                cleaned = clean_manual(raw)
                record_id = f"{topic}({section})" if section else topic

                fout_cleaned.write(
                    json.dumps({"id": record_id, "text": cleaned}, ensure_ascii=False) + '\n'
                )
                cleaned_count += 1

                if fout_train:
                    pairs = generate_training_pairs(cleaned, topic, section)
                    for pair in pairs:
                        fout_train.write(json.dumps(pair, ensure_ascii=False) + '\n')
                        pair_count += 1

                if line_no % 1000 == 0:
                    print(f"🧹 Processed {line_no} lines  |  cleaned: {cleaned_count}",
                          file=sys.stderr, flush=True)

        finally:
            if fout_train:
                fout_train.close()

    print(f"\n✅ Done! Total lines: {line_no}")
    print(f"   Cleaned manuals  → {output_cleaned}  ({cleaned_count} records)")
    if output_training:
        print(f"   Training pairs   → {output_training}  ({pair_count} examples)")

# ----------------------------------------------------------------------
# 4. Training pair generation (unchanged)
# ----------------------------------------------------------------------
def generate_training_pairs(cleaned_text: str, topic: str, section: str) -> list:
    # ... (same as before)
    if not cleaned_text:
        return []
    sections = split_into_sections(cleaned_text)
    if not sections:
        return [make_pair(f"What is the {topic} command?", cleaned_text[:1500])]

    pairs = []
    desc = sections.get('DESCRIPTION') or sections.get('NAME')
    if desc:
        pairs.append(make_pair(f"What does the `{topic}` command do?", desc.strip()))
    syn = sections.get('SYNOPSIS')
    if syn:
        pairs.append(make_pair(f"How do you use `{topic}`?", syn.strip()))
    opts = sections.get('OPTIONS')
    if opts:
        pairs.append(make_pair(f"What are the options of `{topic}`?", opts.strip()))
    ex = sections.get('EXAMPLES')
    if ex:
        pairs.append(make_pair(f"Show me examples of using `{topic}`.", ex.strip()))
    return pairs

def make_pair(user_query: str, assistant_answer: str) -> dict:
    return {
        "messages": [
            {"role": "system", "content": "You are a helpful Linux assistant that explains commands from their man pages."},
            {"role": "user", "content": user_query},
            {"role": "assistant", "content": assistant_answer}
        ]
    }

def split_into_sections(text: str) -> dict:
    sections = {}
    current_heading = None
    current_content = []
    for line in text.split('\n'):
        if re.match(r'^[A-Z][A-Z ]+$', line.strip()) and len(line.strip()) > 2:
            if current_heading:
                sections[current_heading] = '\n'.join(current_content).strip()
            current_heading = line.strip()
            current_content = []
        else:
            if current_heading:
                current_content.append(line)
    if current_heading:
        sections[current_heading] = '\n'.join(current_content).strip()
    return sections

# ----------------------------------------------------------------------
if __name__ == '__main__':
    # You can change these paths
    input_file = "manuals_copy.json"
    cleaned_output = "cleaned_manuals.jsonl"
    training_output = "training_data.jsonl"

    # If you only want the cleaned corpus and no pairs, set training_output = None
    training_output = None

    process_dataset_streaming(input_file, cleaned_output, training_output)