andresnowak/mmlu-auxiliary-train-10-choices

Dataset Card for Augmented MMLU (STEM) with Additional Distractors

Dataset Description

This dataset is an augmented version of the STEM portion of the MMLU auxiliary training set kz919/mmlu-auxiliary-train-auto-labelled, where each original 4-option multiple-choice question has been expanded to include 10 options (A-J) through the addition of six carefully constructed distractors.

Dataset Summary

• Original Dataset: MMLU auxiliary training set (STEM portion)
• Augmentation Method: Added 6 plausible but incorrect options to each question using GPT4o mini
• Total Questions Processed: 13,168
• Success Rate: 99.84% (21 mismatches/failures)
• Final Format: All questions now have 10 answer choices (A-J)

Creation Process

Augmentation Methodology

1. Used GPT4o mini with carefully designed prompt to generate additional distractors
2. Preserved original questions and first 4 answer choices exactly
3. Added 6 new options (E-J) that are:
   • Plausible but incorrect
   • Consistent in style and difficulty
   • Non-redundant with existing options
   • Domain-appropriate

Note: Dataset wasn't checked completely to really see the quality of this distractors few questions where checked and seemed to be good

Prompt Used

### Prompt: Extend MMLU Multiple-Choice Options

You will be given a multiple-choice question from the MMLU dataset, along with four answer choices labeled A through D. Your task is to generate **six additional plausible but incorrect answer choices** (E through J) that are consistent in style, topic, and level of difficulty with the original choices. 

**Do not change the original question or the first four answer choices (A–D)** or their order.

Return the **full list of ten answer choices (A–J)**, keeping the original A–D unchanged, and appending the six new distractors (E–J). Use the exact format below:

A. original_choice_A  
B. original_choice_B  
C. original_choice_C  
D. original_choice_D  
E. new_choice_E  
F. new_choice_F  
G. new_choice_G  
H. new_choice_H  
I. new_choice_I  
J. new_choice_J

**Guidelines for new choices (E–J):**
- Must be plausible but clearly incorrect.
- Should match the grammar and tone of the original options.
- Must not duplicate or closely paraphrase any existing option.
- Should require similar domain knowledge or reasoning as the original options.

**Do not output anything other than the final 10 choices in the specified format.**

---

### Question:
{question}

A. {choice_a}  
B. {choice_b}  
C. {choice_c}  
D. {choice_d}

Model Parameters

{
    "temperature": 0.8,
    "max_tokens": 1000,
    "top_p": 0.8,
    "presence_penalty": 0.1,
    "frequency_penalty": 0.1
}

References of the original dataset cais/mmlu

@article{hendryckstest2021,
      title={Measuring Massive Multitask Language Understanding},
      author={Dan Hendrycks and Collin Burns and Steven Basart and Andy Zou and Mantas Mazeika and Dawn Song and Jacob Steinhardt},
      journal={Proceedings of the International Conference on Learning Representations (ICLR)},
      year={2021}
}

@article{hendrycks2021ethics,
      title={Aligning AI With Shared Human Values},
      author={Dan Hendrycks and Collin Burns and Steven Basart and Andrew Critch and Jerry Li and Dawn Song and Jacob Steinhardt},
      journal={Proceedings of the International Conference on Learning Representations (ICLR)},
      year={2021}
}