src/quality_filter.py

"""
Quality filtering, contamination control, and token-density validation.

Takes the deduplicated documents and applies final quality gates:

  QUALITY FILTERS:
    - Real source code (language signatures at line starts)
    - Balanced delimiters (braces/parens/brackets roughly match)
    - Reasonable length (200–100k chars)
    - Low repetition (unique line ratio)
    - Clean ASCII (low non-ASCII ratio)
    - High code-to-prose ratio
    - Has structure (function/class/struct definitions)

  CONTAMINATION CONTROL:
    - No test files (already filtered in download, double-check here)
    - No auto-generated code markers
    - No license-only files
    - No binary/garbage content
    - No files with extremely high repetition (copy-paste blocks)

  TOKEN DENSITY:
    - Every kept document must be "dense in tokens" — meaning the
      content tokenizes to a meaningful number of tokens relative to
      its character length (no whitespace-padding, no huge comment blocks).
    - Reports token density stats using the project tokenizer.

Input:  data/dedup/documents_dedup.jsonl
Output: data/quality/documents_quality.jsonl  +  quality_stats.json
"""

import json
import os
import re
import sys

PROJECT_DIR = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
DATA_DIR = os.path.join(PROJECT_DIR, "data")
DEDUP_DOCS_PATH = os.path.join(DATA_DIR, "dedup", "documents_dedup.jsonl")
QUALITY_DIR = os.path.join(DATA_DIR, "quality")
QUALITY_DOCS_PATH = os.path.join(QUALITY_DIR, "documents_quality.jsonl")
STATS_PATH = os.path.join(QUALITY_DIR, "quality_stats.json")

# ─── Code structure patterns (must be at line start) ─────────────────────────
CODE_SIGNATURES = {
    "python":   re.compile(r"^(def |class |import |from \S+ import |if __name__|@|async def )", re.M),
    "js_ts":    re.compile(r"^(function |const |let |var |class |export |import |async function|interface |type \w+ =|enum )", re.M),
    "rust":     re.compile(r"^(fn |pub fn |impl |struct |enum |trait |mod |use |pub struct|pub enum|macro_rules!|pub trait)", re.M),
    "go":       re.compile(r"^(func |package |import |type \w+ struct|var |const )", re.M),
    "c":        re.compile(r"^(#include|#define|#ifndef|#ifdef|#if |#endif|typedef |struct \w+|static |void |int |char )", re.M),
    "cpp":      re.compile(r"^(#include|#define|#ifndef|#ifdef|#if |#endif|template |class \w+|namespace |struct \w+|void |int )", re.M),
    "java":     re.compile(r"^(public |private |protected |class \w+|import |package |interface )", re.M),
    "csharp":   re.compile(r"^(public |private |protected |internal |class \w+|using |namespace |interface )", re.M),
    "ruby":     re.compile(r"^(def |class |module |require |require_relative |attr_|include )", re.M),
}

# ─── Contamination / exclusion patterns ──────────────────────────────────────
AUTO_GEN = re.compile(
    r"(?:auto[- ]generated|do not edit|generated by|code generated|"
    r"DO NOT MODIFY|@generated|automatically generated|"
    r"this file was generated)",
    re.IGNORECASE,
)
LICENSE_ONLY = re.compile(r"^(?:/\*|//|#)\s*(?:copyright|licensed|mit license|apache license|bsd license|gnu|gpl)", re.I)
PROSE_LINE = re.compile(r"^[A-Z][a-z]+ .* [a-z]+\.$", re.M)

# ─── Quality scoring ─────────────────────────────────────────────────────────

def detect_language_signatures(doc: dict) -> int:
    """Count code-structure keywords at line starts. Returns count."""
    content = doc["content"]
    lang = doc["language"]
    pat = CODE_SIGNATURES.get(lang)
    if pat is None:
        # Try all
        return sum(len(p.findall(content)) for p in CODE_SIGNATURES.values())
    return len(pat.findall(content))


def score_document(doc: dict) -> tuple[float, str | None]:
    """Score a document 0.0–1.0 on quality. Returns (score, reject_reason)."""
    content = doc["content"]
    lines = content.split("\n")
    n_lines = len(lines)
    length = len(content)

    # ─── Hard rejects (contamination) ──────────────────────────────────────
    if length < 200:
        return 0.0, "too_short"
    if length > 200_000:
        return 0.0, "too_long"
    if n_lines < 5:
        return 0.0, "too_few_lines"

    # Auto-generated code
    if AUTO_GEN.search(content[:2000]):
        return 0.0, "auto_generated"

    # License-only files (first 10 lines are all license comments)
    first_lines = "\n".join(lines[:10])
    if LICENSE_ONLY.search(first_lines) and n_lines < 30:
        return 0.0, "license_only"

    # High non-ASCII (garbage/encoding issues)
    non_ascii = sum(1 for c in content if ord(c) > 127)
    if length > 0 and non_ascii / length > 0.03:
        return 0.0, "high_non_ascii"

    # Extremely high repetition (copy-paste blocks)
    unique_lines = len(set(lines))
    unique_ratio = unique_lines / max(n_lines, 1)
    if unique_ratio < 0.20:
        return 0.0, "high_repetition"

    # ─── Soft scoring ──────────────────────────────────────────────────────
    score = 0.0

    # Base: passes hard filters
    score += 0.15

    # Length quality (sweet spot: 500–30000 chars)
    if 500 <= length <= 30000:
        score += 0.15
    elif 200 <= length <= 80000:
        score += 0.08

    # Delimiter balance
    braces = content.count("{") - content.count("}")
    parens = content.count("(") - content.count(")")
    brackets = content.count("[") - content.count("]")
    total_delims = content.count("{") + content.count("(") + content.count("[")
    if total_delims > 0:
        imbalance = abs(braces) + abs(parens) + abs(brackets)
        balance_ratio = 1.0 - (imbalance / max(total_delims, 1))
        score += 0.15 * max(balance_ratio, 0.0)

    # Code structure density
    struct_count = detect_language_signatures(doc)
    struct_density = min(struct_count / max(n_lines, 1) * 10, 1.0)
    score += 0.15 * struct_density

    # Low repetition (unique line ratio)
    if unique_ratio > 0.7:
        score += 0.10
    elif unique_ratio > 0.5:
        score += 0.05
    else:
        score -= 0.05

    # Clean ASCII
    if length > 0 and non_ascii / length < 0.005:
        score += 0.05

    # Indentation quality (indented lines indicate real code structure)
    indented = sum(1 for l in lines if l.startswith("    ") or l.startswith("\t"))
    if indented > 0 and indented / max(n_lines, 1) > 0.15:
        score += 0.05

    # Penalize high prose ratio (documentation, not code)
    prose_lines = len(PROSE_LINE.findall(content))
    prose_ratio = prose_lines / max(n_lines, 1)
    if prose_ratio > 0.20:
        score -= 0.15

    # Comment density (sweet spot: 3–30%)
    comment_lines = 0
    for line in lines:
        s = line.strip()
        if s.startswith("#") or s.startswith("//") or s.startswith("/*") \
           or s.startswith("*") or s.startswith('"""') or s.startswith("'''") \
           or s.startswith("///") or s.startswith("//!"):
            comment_lines += 1
    comment_ratio = comment_lines / max(n_lines, 1)
    if 0.03 <= comment_ratio <= 0.30:
        score += 0.10
    elif comment_ratio > 0.50:
        score -= 0.10  # too many comments = doc, not code

    return min(max(score, 0.0), 1.0), None


def check_token_density(doc: dict) -> tuple[bool, float]:
    """Check that a document is dense in tokens (not whitespace-padded).

    Returns (passes, chars_per_token_ratio).
    A good code document should have ~2.5-5 chars per token.
    If the ratio is very high (>15), it's likely whitespace/garbage.
    If very low (<1.5), it may be all symbols.
    """
    content = doc["content"]
    # Rough estimate: count non-whitespace characters as a proxy
    # Real tokenization happens with the tokenizer, but this is a fast filter
    non_ws = len(content) - content.count(" ") - content.count("\n") - content.count("\t") - content.count("\r")
    # Approximate token count: split on whitespace + common code delimiters
    approx_tokens = len(re.findall(r"\w+|[^\w\s]", content))
    if approx_tokens == 0:
        return False, 0.0
    chars_per_token = len(content) / approx_tokens
    # Good density: 2.0 - 8.0 chars per token
    passes = 2.0 <= chars_per_token <= 12.0
    return passes, chars_per_token


def main():
    os.makedirs(QUALITY_DIR, exist_ok=True)

    print(f"Loading deduped documents from {DEDUP_DOCS_PATH}...")
    docs = []
    with open(DEDUP_DOCS_PATH, "r", encoding="utf-8") as f:
        for line in f:
            docs.append(json.loads(line))
    print(f"  Loaded {len(docs):,} documents")

    kept = []
    reject_reasons = {}
    scores = []
    token_densities = []

    for i, doc in enumerate(docs):
        score, reject = score_document(doc)

        if reject:
            reject_reasons[reject] = reject_reasons.get(reject, 0) + 1
            continue

        if score < 0.40:
            reject_reasons["low_score"] = reject_reasons.get("low_score", 0) + 1
            continue

        # Token density check
        dense, cpt = check_token_density(doc)
        token_densities.append(cpt)
        if not dense:
            reject_reasons["low_token_density"] = reject_reasons.get("low_token_density", 0) + 1
            continue

        doc = dict(doc)
        doc["quality_score"] = round(score, 4)
        kept.append(doc)
        scores.append(score)

        if (i + 1) % 5000 == 0:
            print(f"  Processed {i+1}/{len(docs)} | kept {len(kept)} | "
                  f"rejected {i+1 - len(kept)}")

    # Stats
    final_size = sum(len(d["content"]) for d in kept)
    avg_score = sum(scores) / len(scores) if scores else 0
    avg_cpt = sum(token_densities) / len(token_densities) if token_densities else 0

    lang_counts = {}
    for d in kept:
        lang_counts[d["language"]] = lang_counts.get(d["language"], 0) + 1

    stats = {
        "input": len(docs),
        "kept": len(kept),
        "rejected": len(docs) - len(kept),
        "reject_reasons": reject_reasons,
        "final_size_mb": final_size / 1e6,
        "avg_quality_score": round(avg_score, 4),
        "avg_chars_per_token": round(avg_cpt, 2),
        "language_distribution": lang_counts,
        "min_score_threshold": 0.40,
    }

    print("\n" + "=" * 60)
    print("QUALITY FILTERING COMPLETE")
    print("=" * 60)
    print(f"  Input:    {len(docs):,}")
    print(f"  Kept:     {len(kept):,}")
    print(f"  Rejected: {len(docs) - len(kept):,}")
    print(f"  Reject reasons: {reject_reasons}")
    print(f"  Final size: {final_size / 1e6:.1f} MB")
    print(f"  Avg quality score: {avg_score:.3f}")
    print(f"  Avg chars/token: {avg_cpt:.2f}")
    print(f"  Languages: {lang_counts}")

    # Write quality-filtered documents
    with open(QUALITY_DOCS_PATH, "w", encoding="utf-8") as f:
        for doc in kept:
            f.write(json.dumps(doc, ensure_ascii=False) + "\n")
    print(f"\nQuality documents written to {QUALITY_DOCS_PATH}")

    with open(STATS_PATH, "w") as f:
        json.dump(stats, f, indent=2)
    print(f"Stats written to {STATS_PATH}")


if __name__ == "__main__":
    main()