feat: normalize code block indentation in training data (#750)

scripts/normalize-code-blocks.py

@@ -0,0 +1,139 @@
+#!/usr/bin/env python3
+"""
+normalize-code-blocks.py — Fix inconsistent indentation in training data code blocks.
+
+When code blocks are embedded in JSONL as triple-quoted strings, indentation
+accumulates from the surrounding context. This script normalizes code block
+content using textwrap.dedent and consistent 4-space indentation.
+
+Usage:
+    python3 scripts/normalize-code-blocks.py training/data/preference_pairs.jsonl
+    python3 scripts/normalize-code-blocks.py --dry-run training/data/*.jsonl
+    python3 scripts/normalize-code-blocks.py --check training/data/*.jsonl  # CI mode
+"""
+
+import argparse
+import json
+import re
+import sys
+import textwrap
+from pathlib import Path
+
+# Matches ```python ... ``` or ``` ... ``` blocks inside string values
+CODE_BLOCK_RE = re.compile(
+    r'(?P<open>```(?:python|py|bash|sh|javascript|js|typescript|ts|go|rust|ruby)?\s*\n)'
+    r'(?P<code>.*?)'
+    r'(?P<close>```)',
+    re.DOTALL,
+)
+
+
+def normalize_code_block(match: re.Match) -> str:
+    """Normalize indentation in a single code block."""
+    open_tag = match.group("open")
+    code = match.group("code")
+    close_tag = match.group("close")
+
+    # Skip empty blocks
+    if not code.strip():
+        return match.group(0)
+
+    # Dedent the code
+    dedented = textwrap.dedent(code)
+
+    # Strip leading/trailing blank lines
+    lines = dedented.split("\n")
+    while lines and not lines[0].strip():
+        lines.pop(0)
+    while lines and not lines[-1].strip():
+        lines.pop()
+
+    normalized = "\n".join(lines)
+
+    return f"{open_tag}{normalized}\n{close_tag}"
+
+
+def process_line(line: str) -> tuple[str, int]:
+    """Process a single JSONL line. Returns (new_line, num_fixes)."""
+    try:
+        obj = json.loads(line)
+    except json.JSONDecodeError:
+        return line, 0
+
+    fixes = 0
+
+    def fix_strings(obj):
+        nonlocal fixes
+        if isinstance(obj, str):
+            original = obj
+            fixed = CODE_BLOCK_RE.sub(normalize_code_block, obj)
+            if fixed != original:
+                fixes += 1
+            return fixed
+        elif isinstance(obj, dict):
+            return {k: fix_strings(v) for k, v in obj.items()}
+        elif isinstance(obj, list):
+            return [fix_strings(item) for item in obj]
+        return obj
+
+    fixed_obj = fix_strings(obj)
+    return json.dumps(fixed_obj, ensure_ascii=False) + "\n", fixes
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Normalize code block indentation in JSONL training data")
+    parser.add_argument("files", nargs="+", help="JSONL files to process")
+    parser.add_argument("--dry-run", action="store_true", help="Show changes without writing")
+    parser.add_argument("--check", action="store_true", help="CI mode: exit 1 if fixes needed")
+    args = parser.parse_args()
+
+    total_fixes = 0
+    total_lines = 0
+    files_changed = 0
+
+    for filepath in args.files:
+        path = Path(filepath)
+        if not path.exists():
+            print(f"SKIP: {path} not found", file=sys.stderr)
+            continue
+
+        lines = path.read_text().splitlines(keepends=True)
+        fixed_lines = []
+        file_fixes = 0
+
+        for i, line in enumerate(lines):
+            if not line.strip():
+                fixed_lines.append(line)
+                continue
+            fixed_line, n = process_line(line)
+            fixed_lines.append(fixed_line)
+            file_fixes += n
+            total_lines += 1
+
+        if file_fixes > 0:
+            files_changed += 1
+            total_fixes += file_fixes
+            print(f"{'CHECK' if args.check else 'FIX'}: {path} — {file_fixes} code blocks normalized")
+
+            if args.check:
+                # Show diff
+                for i, (old, new) in enumerate(zip(lines, fixed_lines)):
+                    if old != new:
+                        print(f"  Line {i+1}: indentation changed")
+            elif not args.dry_run:
+                path.write_text("".join(fixed_lines))
+                print(f"  Written: {path}")
+        else:
+            print(f"OK: {path} — no indentation issues")
+
+    print(f"\nSummary: {total_fixes} code blocks fixed across {files_changed} files ({total_lines} lines processed)")
+
+    if args.check and total_fixes > 0:
+        print("FAIL: Code block indentation issues found. Run without --check to fix.")
+        sys.exit(1)
+
+    sys.exit(0)
+
+
+if __name__ == "__main__":
+    main()