scripts/pipeline.py

"""
MiniF2F Alpha-Renaming Pipeline

This script creates an alpha-renamed version of the MiniF2F test dataset.
Alpha-renaming replaces variable names with Greek letters (α, β, γ, etc.)
using Lean 4 meta-programming to ensure syntactic correctness.

Pipeline steps:
  1. Download MiniF2F from HuggingFace
  2. Preprocess: remove comments, extract propositions
  3. Alpha-rename using Lean 4 meta-programming (AlphaRenaming.lean)
  4. Verify with kimina-lean-server (check syntax validity)
  5. Upload verified dataset to HuggingFace

Usage:
  uv run python pipeline.py --repo-id ChristianZ97/minif2f_test-AlphaRenamed
"""

import os
import re
import sys
import json
import tempfile
import subprocess
import argparse
from typing import List, Dict, Optional
from dataclasses import dataclass, field
from datetime import datetime

from tqdm import tqdm
from datasets import Dataset, load_dataset
from huggingface_hub import login, DatasetCard

# Add parent directory to path for kimina_client
sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))

try:
    from kimina_client import KiminaClient
except ImportError:
    KiminaClient = None
    print("Warning: kimina_client not found. Install or check path.")


# ============================================================================
# Configuration
# ============================================================================

PROJECT_DIR = os.path.dirname(os.path.abspath(__file__))

# Standard Lean 4 header for Mathlib
DEFAULT_LEAN_HEADER = """import Mathlib
import Aesop

set_option maxHeartbeats 0

open BigOperators Real Nat Topology Rat
"""


@dataclass
class LeanState:
    """Result of Lean verification."""
    status: str  # "Success", "Error", "Sorry"
    is_syntax_correct: bool
    is_proven: bool
    has_sorry: bool
    error_messages: List[str] = field(default_factory=list)


@dataclass
class Problem:
    """A MiniF2F problem through the pipeline."""
    unique_id: str
    original_statement: str      # Raw formal_statement from HF
    cleaned_statement: str       # After removing comments
    original_prop: str           # Extracted proposition (∀ form)
    renamed_prop: Optional[str]  # After alpha-renaming
    verified: bool               # Passed kimina verification
    informal_statement: str      # Natural language description
    error_message: Optional[str] = None


# ============================================================================
# Step 1: Download MiniF2F
# ============================================================================

def download_minif2f() -> List[Dict]:
    """Download MiniF2F test dataset from HuggingFace (AI-MO/minif2f_test)."""
    print("=" * 60)
    print("Step 1: Downloading MiniF2F from HuggingFace")
    print("=" * 60)
    
    dataset = load_dataset("AI-MO/minif2f_test", split="train")
    problems = list(dataset)
    print(f"  Downloaded {len(problems)} problems")
    return problems


# ============================================================================
# Step 2: Preprocess
# ============================================================================

def preprocess_statement(formal_statement: str) -> str:
    """Remove comments and clean up formal statement."""
    # Remove block comments /-- ... -/
    cleaned = re.sub(r'/--[\s\S]*?-/', '', formal_statement)
    # Remove line comments -- ...
    cleaned = re.sub(r'--.*', '', cleaned)
    # Clean up whitespace
    cleaned = re.sub(r'\n\s*\n', '\n', cleaned)
    return cleaned.strip()


def extract_proposition(cleaned_statement: str) -> Optional[str]:
    """
    Extract proposition from theorem declaration.
    Example: "theorem foo (x : ℕ) : x = x := by sorry" -> "(∀ (x : ℕ), x = x)"
    """
    # Match: theorem name params : goal := ...
    pattern = r'theorem\s+\w+\s*(.*?)\s*:=\s*(?:by)?'
    match = re.search(pattern, cleaned_statement, re.DOTALL)
    
    if not match:
        return None
    
    signature = match.group(1).strip()
    
    # Find outermost : separating params from goal
    depth = 0
    colon_pos = -1
    for i, c in enumerate(signature):
        if c in '([{':
            depth += 1
        elif c in ')]}':
            depth -= 1
        elif c == ':' and depth == 0:
            colon_pos = i
    
    if colon_pos == -1:
        return None
    
    params = signature[:colon_pos].strip()
    goal = signature[colon_pos + 1:].strip()
    
    if params:
        return f"(∀ {params}, {goal})"
    else:
        return f"({goal})"


def preprocess_problems(raw_problems: List[Dict]) -> List[Problem]:
    """Preprocess all problems: clean and extract propositions."""
    print("\n" + "=" * 60)
    print("Step 2: Preprocessing")
    print("=" * 60)
    
    problems = []
    failed = 0
    
    for sample in raw_problems:
        unique_id = sample.get("name", "unknown")
        formal_stmt = sample.get("formal_statement", "")
        informal = sample.get("informal_prefix", "")
        
        # Clean informal statement
        if informal.startswith("/- "):
            informal = informal[3:]
        if informal.endswith(" -/"):
            informal = informal[:-3]
        
        cleaned = preprocess_statement(formal_stmt)
        prop = extract_proposition(cleaned)
        
        if prop:
            problems.append(Problem(
                unique_id=unique_id,
                original_statement=formal_stmt,
                cleaned_statement=cleaned,
                original_prop=prop,
                renamed_prop=None,
                verified=False,
                informal_statement=informal.strip(),
            ))
        else:
            failed += 1
    
    print(f"  Extracted: {len(problems)} propositions")
    print(f"  Failed: {failed}")
    return problems


# ============================================================================
# Step 3: Alpha-Renaming (Lean Meta-programming)
# ============================================================================

def create_lean_file_for_batch(problems: List[Problem]) -> str:
    """Create Lean file using #alpha_rename_id command for batch processing."""
    lines = [
        "import AlphaRenaming",
        "open AlphaRenaming",
        "",
    ]
    
    for p in problems:
        safe_id = re.sub(r'[^a-zA-Z0-9_]', '_', p.unique_id)
        lines.append(f"#alpha_rename_id [{safe_id}] {p.original_prop}")
        lines.append("")
    
    return "\n".join(lines)


def parse_lean_output(output: str, problems: List[Problem]) -> Dict[str, str]:
    """Parse Lean output to extract renamed propositions."""
    renamed = {}
    
    # Build safe_id -> unique_id map
    id_map = {}
    for p in problems:
        safe_id = re.sub(r'[^a-zA-Z0-9_]', '_', p.unique_id)
        id_map[safe_id] = p.unique_id
    
    # Parse [RENAMED:id]...[/RENAMED:id] markers
    pattern = r'\[RENAMED:(\w+)\]\s*(.*?)\s*\[/RENAMED:\1\]'
    for match in re.finditer(pattern, output, re.DOTALL):
        safe_id = match.group(1)
        renamed_expr = match.group(2).strip()
        if safe_id in id_map:
            renamed[id_map[safe_id]] = renamed_expr
    
    return renamed


def alpha_rename_batch(problems: List[Problem], batch_size: int = 30) -> None:
    """Alpha-rename using Lean meta-programming in batches."""
    print("\n" + "=" * 60)
    print("Step 3: Alpha-Renaming (Lean Meta-programming)")
    print("=" * 60)
    
    total_renamed = 0
    
    for i in tqdm(range(0, len(problems), batch_size), desc="Renaming"):
        batch = problems[i:i + batch_size]
        lean_code = create_lean_file_for_batch(batch)
        
        with tempfile.NamedTemporaryFile(
            mode='w', suffix='.lean', dir=PROJECT_DIR, delete=False
        ) as f:
            f.write(lean_code)
            temp_file = f.name
        
        try:
            result = subprocess.run(
                ['lake', 'env', 'lean', temp_file],
                cwd=PROJECT_DIR,
                capture_output=True,
                text=True,
                timeout=120
            )
            
            output = result.stdout + result.stderr
            renamed = parse_lean_output(output, batch)
            
            for p in batch:
                if p.unique_id in renamed:
                    p.renamed_prop = renamed[p.unique_id]
                    total_renamed += 1
                    
        except subprocess.TimeoutExpired:
            print(f"  Warning: Batch {i//batch_size} timed out")
        except Exception as e:
            print(f"  Warning: Batch {i//batch_size} error: {e}")
        finally:
            if os.path.exists(temp_file):
                os.remove(temp_file)
    
    print(f"  Renamed: {total_renamed}/{len(problems)}")


# ============================================================================
# Step 4: Verify with kimina-lean-server
# ============================================================================

def get_val(obj, key: str, default=None):
    """Safely get value from dict or object."""
    if isinstance(obj, dict):
        return obj.get(key, default)
    return getattr(obj, key, default)


def parse_lean_response(response) -> LeanState:
    """Parse KiminaClient response."""
    has_error = False
    has_sorry = False
    error_msgs = []

    if not response or not hasattr(response, 'results') or not response.results:
        return LeanState("Error", False, False, False, ["No results"])

    result = response.results[0]
    lean_resp = getattr(result, "response", {})

    messages = get_val(lean_resp, "messages", []) or []
    for msg in messages:
        severity = get_val(msg, "severity", "")
        data = get_val(msg, "data", "")

        if severity == "error":
            has_error = True
            error_msgs.append(data)

        if "declaration uses 'sorry'" in str(data):
            has_sorry = True

    sorries = get_val(lean_resp, "sorries", []) or []
    if sorries:
        has_sorry = True

    if has_error:
        return LeanState("Error", False, False, has_sorry, error_msgs)
    elif has_sorry:
        return LeanState("Sorry", True, False, True)
    else:
        return LeanState("Success", True, True, False)


def verify_single(client, problem: Problem) -> None:
    """Verify single problem using kimina-lean-server (append sorry)."""
    prop_to_verify = problem.renamed_prop or problem.original_prop
    
    if not prop_to_verify:
        return
    
    safe_name = re.sub(r'[^a-zA-Z0-9_]', '_', problem.unique_id)
    code = DEFAULT_LEAN_HEADER + f"\ntheorem {safe_name} : {prop_to_verify} := by\n  sorry\n"
    
    try:
        response = client.check(code)
        state = parse_lean_response(response)
        problem.verified = state.status in ["Sorry", "Success"]
        
        if state.error_messages:
            problem.error_message = state.error_messages[0][:300]
    except Exception as e:
        problem.error_message = str(e)[:300]


def verify_all(problems: List[Problem]) -> None:
    """Verify all problems with kimina-lean-server."""
    print("\n" + "=" * 60)
    print("Step 4: Verifying with kimina-lean-server")
    print("=" * 60)
    
    if KiminaClient is None:
        print("  Error: kimina_client not available")
        return
    
    try:
        client = KiminaClient()
    except Exception as e:
        print(f"  Error initializing client: {e}")
        return
    
    to_verify = [p for p in problems if p.renamed_prop or p.original_prop]
    
    for p in tqdm(to_verify, desc="Verifying"):
        verify_single(client, p)
    
    verified_count = sum(1 for p in problems if p.verified)
    error_count = sum(1 for p in to_verify if not p.verified)
    
    print(f"  Verified OK: {verified_count}")
    print(f"  Errors: {error_count}")


# ============================================================================
# Step 5: Upload to HuggingFace
# ============================================================================

def create_hf_dataset(problems: List[Problem]) -> Dataset:
    """Create HuggingFace Dataset from ALL processed problems."""
    data = {
        "unique_id": [],
        "informal_statement": [],
        "original_statement": [],
        "renamed_statement": [],
        "original_prop": [],
        "renamed_prop": [],
        "kimina_verified": [],
        "error_message": [],
    }
    
    for p in problems:
        # Build renamed_statement if renamed_prop exists (use anonymous name)
        if p.renamed_prop:
            full_theorem = DEFAULT_LEAN_HEADER + f"\ntheorem anonymous : {p.renamed_prop} := by\n  sorry"
        else:
            full_theorem = ""  # Empty if renaming failed
        
        data["unique_id"].append(p.unique_id)
        data["informal_statement"].append(p.informal_statement)
        data["original_statement"].append(p.original_statement)
        data["renamed_statement"].append(full_theorem)
        data["original_prop"].append(p.original_prop)
        data["renamed_prop"].append(p.renamed_prop or "")
        data["kimina_verified"].append(p.verified)
        data["error_message"].append(p.error_message or "")
    
    return Dataset.from_dict(data)


def upload_to_hf(dataset: Dataset, repo_id: str, private: bool = False) -> None:
    """Upload dataset and README to HuggingFace Hub."""
    print("\n" + "=" * 60)
    print("Step 5: Uploading to HuggingFace")
    print("=" * 60)
    
    print(f"  Repo: {repo_id}")
    print(f"  Samples: {len(dataset)}")
    
    # Count stats
    verified_count = sum(1 for v in dataset["kimina_verified"] if v)
    renamed_count = sum(1 for r in dataset["renamed_prop"] if r)
    
    commit_msg = f"Upload alpha-renamed MiniF2F ({datetime.now().strftime('%Y-%m-%d %H:%M')})"
    
    # Push dataset
    dataset.push_to_hub(
        repo_id,
        private=private,
        commit_message=commit_msg,
    )
    
    # Upload README
    readme_path = os.path.join(PROJECT_DIR, "README.md")
    if os.path.exists(readme_path):
        with open(readme_path, 'r') as f:
            readme_content = f.read()
        
        # Create dataset card with YAML front matter
        yaml_header = """---
language:
- en
license: mit
size_categories:
- n<1K
task_categories:
- text-generation
tags:
- lean4
- mathlib
- theorem-proving
- alpha-renaming
- minif2f
---

"""
        full_readme = yaml_header + readme_content
        card = DatasetCard(full_readme)
        card.push_to_hub(repo_id)
        print(f"  ✓ README uploaded")
    
    print(f"  ✓ Uploaded to https://huggingface.co/datasets/{repo_id}")
    print(f"  Stats: {renamed_count} renamed, {verified_count} verified, {len(dataset)} total")


# ============================================================================
# Main Pipeline
# ============================================================================

def run_pipeline(
    repo_id: str,
    limit: Optional[int] = None,
    skip_verify: bool = False,
    skip_upload: bool = False,
    include_unverified: bool = False,
    private: bool = False,
    save_local: Optional[str] = None,
    hf_token: Optional[str] = None,
):
    """Run complete alpha-renaming pipeline."""
    print("\n" + "=" * 60)
    print("  MiniF2F Alpha-Renaming Pipeline")
    print("=" * 60)
    
    # HuggingFace login
    token = hf_token or os.environ.get("HF_TOKEN")
    if token:
        login(token=token)
    
    # Step 1: Download
    raw_problems = download_minif2f()
    if limit:
        raw_problems = raw_problems[:limit]
    
    # Step 2: Preprocess
    problems = preprocess_problems(raw_problems)
    
    # Step 3: Alpha-rename
    alpha_rename_batch(problems)
    
    # Step 4: Verify
    if not skip_verify:
        verify_all(problems)
    else:
        for p in problems:
            if p.renamed_prop:
                p.verified = True
    
    # Save local copy
    if save_local:
        local_data = [
            {
                "unique_id": p.unique_id,
                "original_prop": p.original_prop,
                "renamed_prop": p.renamed_prop,
                "verified": p.verified,
                "error": p.error_message,
            }
            for p in problems
        ]
        with open(save_local, 'w') as f:
            json.dump(local_data, f, indent=2, ensure_ascii=False)
        print(f"\nSaved local copy to {save_local}")
    
    # Step 5: Upload
    if not skip_upload:
        dataset = create_hf_dataset(problems)  # Include ALL problems
        
        if len(dataset) == 0:
            print("\nError: No valid samples to upload")
            return
        
        upload_to_hf(dataset, repo_id, private=private)
    else:
        print("\n(Skipping HuggingFace upload)")
    
    # Summary
    print("\n" + "=" * 60)
    print("  Summary")
    print("=" * 60)
    total = len(problems)
    renamed = sum(1 for p in problems if p.renamed_prop)
    verified = sum(1 for p in problems if p.verified)
    print(f"  Total: {total}")
    print(f"  Renamed: {renamed} ({100*renamed/total:.1f}%)")
    print(f"  Verified: {verified} ({100*verified/total:.1f}%)")


# ============================================================================
# CLI
# ============================================================================

def main():
    parser = argparse.ArgumentParser(description="MiniF2F Alpha-Renaming Pipeline")
    parser.add_argument("--repo-id", "-r", type=str, required=True,
                        help="HuggingFace repo ID")
    parser.add_argument("--limit", "-n", type=int, default=None,
                        help="Limit number of problems (for testing)")
    parser.add_argument("--skip-verify", action="store_true",
                        help="Skip kimina-lean-server verification")
    parser.add_argument("--skip-upload", action="store_true",
                        help="Skip HuggingFace upload")
    parser.add_argument("--include-unverified", action="store_true",
                        help="Include unverified problems")
    parser.add_argument("--private", action="store_true",
                        help="Make dataset private")
    parser.add_argument("--save-local", "-o", type=str, default=None,
                        help="Save to local JSON file")
    parser.add_argument("--token", type=str, default=None,
                        help="HuggingFace token")
    
    args = parser.parse_args()
    
    run_pipeline(
        repo_id=args.repo_id,
        limit=args.limit,
        skip_verify=args.skip_verify,
        skip_upload=args.skip_upload,
        include_unverified=args.include_unverified,
        private=args.private,
        save_local=args.save_local,
        hf_token=args.token,
    )


if __name__ == "__main__":
    main()