scripts/shuffle_dataset.py

#!/usr/bin/env python3
"""
Shuffle MCQ candidates across all GomParam-v1 JSON modules.

This script randomizes the position of the correct answer in every question
to eliminate positional bias (where ~88% of correct answers were at index 0).
Uses a fixed seed for reproducibility.
"""

import json
import random
from pathlib import Path

SEED = 42
DATA_DIR = Path(__file__).resolve().parent.parent / "data"

random.seed(SEED)

stats_before = {0: 0, 1: 0, 2: 0, 3: 0}
stats_after  = {0: 0, 1: 0, 2: 0, 3: 0}
total_items  = 0

for json_file in sorted(DATA_DIR.glob("*.json")):
    with open(json_file, "r", encoding="utf-8") as f:
        items = json.load(f)

    modified = False
    for item in items:
        candidates = item.get("candidates", [])
        correct_idx = item.get("correct", -1)

        if not candidates or correct_idx < 0 or correct_idx >= len(candidates):
            continue

        total_items += 1
        stats_before[correct_idx] = stats_before.get(correct_idx, 0) + 1

        # Build index list and shuffle
        correct_answer = candidates[correct_idx]
        indices = list(range(len(candidates)))
        random.shuffle(indices)

        # Rearrange candidates and find new correct index
        new_candidates = [candidates[i] for i in indices]
        new_correct = indices.index(correct_idx)

        item["candidates"] = new_candidates
        item["correct"] = new_correct
        modified = True

        stats_after[new_correct] = stats_after.get(new_correct, 0) + 1

    if modified:
        with open(json_file, "w", encoding="utf-8") as f:
            json.dump(items, f, ensure_ascii=False, indent=2)
        print(f"  ✓ {json_file.name:40s} {len(items)} items shuffled")

print(f"\nTotal items processed: {total_items}")
print(f"\nBEFORE shuffle — correct answer position distribution:")
for k, v in sorted(stats_before.items()):
    print(f"  Index {k}: {v:5d}  ({v/total_items*100:.1f}%)")

print(f"\nAFTER shuffle — correct answer position distribution:")
for k, v in sorted(stats_after.items()):
    print(f"  Index {k}: {v:5d}  ({v/total_items*100:.1f}%)")