experiments/learnability/split_examples.py

#!/usr/bin/env python3
"""
Step 3: Extract training examples for a specific filter condition.

Takes a single filtered_candidates.json and the full examples_grouped.json
(from prove_and_extract.py on the 'none' filter), looks up proof results
for each candidate, and writes examples.json.

Optionally subsample N candidates (randomly) from the filtered set.

Usage:
    python split_examples.py \
        --grouped .../none/examples_grouped.json \
        --candidates .../novel-1.0/filtered_candidates.json \
        --output_dir .../novel-1.0

    # Keep only 50 randomly sampled candidates:
    python split_examples.py \
        --grouped .../none/examples_grouped.json \
        --candidates .../novel-1.0/filtered_candidates.json \
        --output_dir .../novel-1.0-N50 \
        -N 50 --seed 42
"""

import argparse
import json
import random
from pathlib import Path


def main():
    parser = argparse.ArgumentParser(
        description='Extract training examples for a filtered candidate set')
    parser.add_argument('--grouped', required=True,
                        help='Path to examples_grouped.json (from prove_and_extract.py on none)')
    parser.add_argument('--candidates', required=True,
                        help='Path to filtered_candidates.json')
    parser.add_argument('--output_dir', required=True,
                        help='Output directory for examples')
    parser.add_argument('-N', type=int, default=None,
                        help='Randomly sample N candidates (default: keep all)')
    parser.add_argument('--seed', type=int, default=42,
                        help='Random seed for subsampling (default: 42)')
    args = parser.parse_args()

    # Load grouped examples (keyed by candidate statement)
    print(f"Loading grouped examples from {args.grouped}")
    with open(args.grouped) as f:
        grouped = json.load(f)

    stmt_to_entry = {}
    for entry in grouped['problems']:
        stmt_to_entry[entry['candidate']] = entry
    print(f"  {len(stmt_to_entry)} candidates with proof results")

    # Load filtered candidates
    print(f"\nLoading candidates from {args.candidates}")
    with open(args.candidates) as f:
        cand_data = json.load(f)

    filter_name = cand_data['filter']
    candidates = cand_data['candidates']
    print(f"  Filter: {filter_name}")
    print(f"  Candidates: {len(candidates)}")

    # Subsample if requested
    if args.N is not None and args.N < len(candidates):
        rng = random.Random(args.seed)
        candidates = rng.sample(candidates, args.N)
        print(f"  Subsampled to N={args.N} (seed={args.seed})")

    # Look up examples
    problems = []
    flat_examples = []
    missing = 0

    for stmt in candidates:
        entry = stmt_to_entry.get(stmt)
        if entry is None:
            missing += 1
            continue
        problems.append(entry)
        flat_examples.extend(entry['examples'])

    n_proved = sum(1 for p in problems if p['success'])
    n_examples = len(flat_examples)
    n_per_proved = [len(p['examples']) for p in problems if p['success']]
    avg_ex = sum(n_per_proved) / len(n_per_proved) if n_per_proved else 0

    print(f"\n  Matched: {len(problems)}, Missing: {missing}")
    print(f"  Proved: {n_proved}, Examples: {n_examples}, Avg/proved: {avg_ex:.1f}")

    # Save
    out_dir = Path(args.output_dir)
    out_dir.mkdir(parents=True, exist_ok=True)

    with open(out_dir / 'examples_grouped.json', 'w') as f:
        json.dump({
            'filter': filter_name,
            'source_grouped': str(args.grouped),
            'source_candidates': str(args.candidates),
            'n_candidates': len(candidates),
            'n_sampled': args.N,
            'sample_seed': args.seed if args.N else None,
            'n_proved': n_proved,
            'n_examples': n_examples,
            'problems': problems,
        }, f, indent=2)

    with open(out_dir / 'examples.json', 'w') as f:
        json.dump(flat_examples, f)

    with open(out_dir / 'candidates.txt', 'w') as f:
        for stmt in candidates:
            f.write(stmt + '\n')

    summary = {
        'filter': filter_name,
        'source_grouped': str(args.grouped),
        'source_candidates': str(args.candidates),
        'n_candidates': len(candidates),
        'n_sampled': args.N,
        'n_proved': n_proved,
        'n_examples': n_examples,
        'avg_examples_per_problem': avg_ex,
    }
    with open(out_dir / 'summary.json', 'w') as f:
        json.dump(summary, f, indent=2)

    print(f"\nSaved to {out_dir}/")
    print(f"  examples.json ({n_examples} examples)")
    print(f"  examples_grouped.json ({len(problems)} problems)")
    print(f"  candidates.txt ({len(candidates)} statements)")
    print(f"  summary.json")


if __name__ == '__main__':
    main()