File size: 5,227 Bytes
58c635d
22fb00d
 
 
 
 
 
 
 
 
 
 
 
 
 
58c635d
22fb00d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
aa8b00a
22fb00d
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
---
annotations_creators:
- expert-generated
language:
- en
license: mit
multilinguality:
- monolingual
size_categories:
- 10K<n<100K
task_categories:
- multiple-choice
task_ids:
- multiple-choice-qa
pretty_name: MNLP M3 MCQA Dataset
---

# MNLP M3 MCQA Dataset

The **MNLP M3 MCQA Dataset** is a carefully curated collection of **Multiple-Choice Question Answering (MCQA)** examples, unified from several academic and benchmark datasets.

Developed as part of the *CS-552: Modern NLP* course at EPFL (Spring 2025), this dataset is designed for training and evaluating models on multiple-choice QA tasks, particularly in the **STEM** and general knowledge domains.

## Key Features

- ~30,000 MCQA questions
- 6 diverse sources: `SciQ`, `OpenBookQA`, `MathQA`, `ARC-Easy`, `ARC-Challenge`, and `MedMCQA`
- Each question has exactly 4 options (A–D) and one correct answer
- Covers a wide range of topics: science, technology, engineering, mathematics, and general knowledge

## Dataset Structure

Each example is a dictionary with the following fields:

| Field     | Type     | Description                                       |
|-----------|----------|---------------------------------------------------|
| `dataset` | `string` | Source dataset (`sciq`, `openbookqa`, etc.)      |
| `id`      | `string` | Unique identifier for the question                |
| `question`| `string` | The question text                                 |
| `choices` | `list`   | List of 4 answer options (corresponding to A–D)   |
| `answer`  | `string` | The correct option, as a letter: `"A"`, `"B"`, `"C"`, or `"D"` |
| `support` | `string` | A brief explanation or fact supporting the correct answer when available |

```markdown
Example:
```json
{
  "dataset": "sciq",
  "id": "sciq_01_00042",
  "question": "What does a seismograph measure?",
  "choices": ["Earthquakes", "Rainfall", "Sunlight", "Temperature"],
  "answer": "A",
  "support": "A seismograph is an instrument that detects and records earthquakes."
}
```

## Source Datasets

This dataset combines multiple high-quality MCQA sources to support research and fine-tuning in STEM education and reasoning. The full corpus contains **29,870 multiple-choice questions** from the following sources:

| Source (Hugging Face)                       | Name                | Size   | Description & Role in the Dataset                                                                                                                                                                                             |
| ------------------------------------------- | ------------------- | ------ | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| `allenai/sciq`                              | **SciQ**            | 11,679 | **Science questions** (Physics, Chemistry, Biology, Earth science). Crowdsourced with 4 answer choices and optional supporting evidence. Used to provide **well-balanced, factual STEM questions** at a middle/high-school level. |
| `allenai/openbookqa`                        | **OpenBookQA**      | 4,957  | Science exam-style questions requiring **multi-step reasoning** and use of **commonsense or external knowledge**. Contributes more **challenging** and **inference-based** questions.                                         |
| `allenai/math_qa`                           | **MathQA**          | 5,000  | Subsample of quantitative math word problems derived from AQuA-RAT, annotated with structured answer options. Introduces **numerical reasoning** and **problem-solving** components into the dataset.                         |
| `allenai/ai2_arc` (config: `ARC-Easy`)      | **ARC-Easy**        | 2,140  | Science questions at the middle school level. Useful for testing **basic STEM understanding** and **factual recall**. Filtered to retain only valid 4-choice entries.                                                         |
| `allenai/ai2_arc` (config: `ARC-Challenge`) | **ARC-Challenge**   | 1,094  | More difficult science questions requiring **reasoning and inference**. Widely used as a benchmark for evaluating LLMs. Also filtered for clean MCQA format compatibility.                                                    |
| `openlifescienceai/medmcqa`                | **MedMCQA**         | 5,000  | A subsample of multiple-choice questions on **medical topics** from various exams, filtered for a single-choice format. Contains real-world and domain-specific **clinical reasoning** questions covering various medical disciplines.         |

## Intended Applications and Structure

This dataset is split into three parts: 

- `train` (~85%) — for training MCQA models
- `validation` (~15%) — for tuning and monitoring performance during training

It is suitable for multiple-choice question answering tasks, especially in the **STEM** domain (Science, Technology, Engineering, Mathematics).  

## Author

This dataset was created and published by [Youssef Belghmi](https://huggingface.co/youssefbelghmi) as part of the *CS-552: Modern NLP* course at EPFL (Spring 2025).