Rijgersberg/dutch-central-exam-mcq-shuffled-answers

Multiple Choice Questions of the Dutch Central Exam 1999-2024 (with the answer options shuffled)

This is reproduction of jjzha/dutch-central-exam-mcq with the answer options shuffled. Be sure to credit the original dataset when using this modified version.

Distribution of correct answer options.

The original dataset has an uneven distribution of correct answer options:

Answer	train	dev	test
1	56	—	16
2	447	2	128
3	457	6	115
4	215	3	114
5	20	—	11
6	11	—	3

The lack of questions with the correct answer "1" (A) is due to a bug in the data processing. When training on this set, a model might first fit the underlying answer distribution instead of learning the task.

This version of the dataset shuffles the answer options in order to regain a more equal distribution:

Answer	train	dev	test
1	303	5	92
2	309	6	102
3	360	—	83
4	182	—	97
5	32	—	7
6	18	—	6
7	2	—	—

The fact that there are still fewer questions with correct answers "4", "5", "6" and "7" is caused by the fact that some questions have fewer answer options than others.

Code

The following code was used to created the shuffled version:

import random
from collections import Counter

from datasets import load_dataset


random.seed(42)

def count(dataset):
    for split in dataset:
        answers = dataset[split]['answer']
        counter = Counter(answers)

        print(split)
        for i in sorted(counter):
            print(f'{i}: {counter[i]}')

def shuffle_answers(options, answer):
    """Shuffle the answer options to obtain a uniform distribution of answers options.
    
    This is necessary because the original dataset is biased against answer "1" ()"""
    correct_answer = options[int(answer) - 1]

    random.shuffle(options)

    new_answer_0index = options.index(correct_answer)

    # original dataset has the 1-indexed answer number as a string
    new_answer = str(new_answer_0index + 1)

    return {
        'options': options,
        'answer': new_answer
    }

dataset = load_dataset('jjzha/dutch-central-exam-mcq')

print('original distribution:')
count(dataset)

shuffled_dataset = dataset.map(shuffle_answers, input_columns=['options', 'answer'],
                               load_from_cache_file=False)

print('\nshuffled distribution:')
count(shuffled_dataset)

# check that answers are the same
for split in dataset:
    for orig_row, shuffled_row in zip(dataset[split], shuffled_dataset[split], strict=True):
        orig_ans = orig_row['options'][int(orig_row['answer']) - 1]
        shuffled_ans = shuffled_row['options'][int(shuffled_row['answer']) - 1]

        if orig_ans != shuffled_ans:
            raise ValueError(f'{orig_ans} is not the same as {shuffled_ans}')

shuffled_dataset.push_to_hub('Rijgersberg/dutch-central-exam-mcq-shuffled-answers')