README.md

---
dataset_info:
  features:
  - name: question
    dtype: string
  - name: options
    sequence: string
  - name: rationale
    dtype: string
  - name: label
    dtype: string
  - name: label_idx
    dtype: int64
  - name: dataset
    dtype: string
  splits:
  - name: train
    num_bytes: 203046319
    num_examples: 200000
  - name: validation
    num_bytes: 264310
    num_examples: 519
  download_size: 122985245
  dataset_size: 203310629
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: validation
    path: data/validation-*
license: apache-2.0
task_categories:
- multiple-choice
language:
- en
size_categories:
- 100K<n<1M
---


# MNLP M2 MCQA Dataset

A unified multiple-choice question answering (MCQA) benchmark on STEM subjects combining samples from OpenBookQA, SciQ, MMLU-auxiliary, AQUA-Rat, and MedMCQA.

## Dataset Summary

This dataset merges five existing science and knowledge-based MCQA datasets into one standardized format:

| Source | Train samples |
| ---------- | ------------: |
| OpenBookQA | 4 900 |
| SciQ | 10 000 |
| MMLU-aux | 85 100 |
| AQUA-Rat | 50 000 |
| MedMCQA | 50 000 |
| **Total** | **200 000** |

## Supported Tasks and Leaderboards

* **Task:** Multiple-Choice Question Answering (`multiple-choice-question-answering`)
* **Metrics:** Accuracy

## Languages

* English

## Dataset Structure

Each example has the following fields:

| Name | Type | Description |
| ----------- | -------------- | ------------------------------------------------ |
| `question` | `string` | The question stem. |
| `options` | `list[string]` | List of 4-5 answer choices. |
| `label` | `string` | The correct answer letter, e.g. `"A"`, or `"a"`. |
| `label_idx` | `int` | Zero-based index of the correct answer (0–4). |
| `rationale` | `string` | (Optional) Supporting fact or rationale text. |
| `dataset` | `string` | Source dataset name (`openbookqa`, `sciq`, etc.) |

### Splits

```
DatasetDict({
train: Dataset(num_rows=200000),
validation: Dataset(num_rows=519),
})
```

## Dataset Creation

1. **Source Datasets**

* OpenBookQA (`allenai/openbookqa`)
* SciQ (`allenai/sciq`)
* MMLU-auxiliary (`cais/mmlu`, config=`all`)
* AQUA-Rat (`deepmind/aqua_rat`)
* MedMCQA (`openlifescienceai/medmcqa`)

2. **Sampling**
We sample each training split down to a fixed size (4 900–85 100 examples). Validation examples are sampled per source by first computing each dataset’s original validation-to-train ratio (len(validation)/len(train)), taking the minimum of these ratios and 5 %, and then holding out that fraction from each source.

3. **Unification**
All examples are mapped to a common schema (`question`, `options`, `label`, …) with minimal preprocessing.

4. **Push to Hub**

```python
from datasets import DatasetDict, load_dataset, concatenate_datasets

# after loading, sampling, mapping…
ds = DatasetDict({"train": combined, "validation": val_combined})
ds.push_to_hub("NicoHelemon/MNLP_M2_mcqa_dataset", private=False)
```

## Usage

```python
from datasets import load_dataset

ds = load_dataset("NicoHelemon/MNLP_M2_mcqa_dataset")
print(ds["train"][0])
# {
# "question": "What can genes do?",
# "options": ["Give a young goat hair that looks like its mother's hair", ...],
# "label": "A",
# "label_idx": 0,
# "rationale": "Key fact: genes are a vehicle for passing inherited…",
# "dataset": "openbookqa"
# }
```

## Licensing

This collection is released under the **Apache-2.0** license.
Original source datasets may carry their own licenses—please cite appropriately.

## Citation

If you use this dataset, please cite:

```bibtex


@misc
{helemon2025m2mcqa,
title = {MNLP M2 MCQA Dataset},
author = {Nicolas Gonzalez},
year = 2025,
howpublished = {\url{https://huggingface.co/datasets/NicoHelemon/MNLP_M2_mcqa_dataset}},
}
```

And please also cite the original datasets:

```bibtex

@misc{mihaylov2018suitarmorconductelectricity,
      title={Can a Suit of Armor Conduct Electricity? A New Dataset for Open Book Question Answering}, 
      author={Todor Mihaylov and Peter Clark and Tushar Khot and Ashish Sabharwal},
      year={2018},
      eprint={1809.02789},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/1809.02789}, 
}

@misc{welbl2017crowdsourcingmultiplechoicescience,
      title={Crowdsourcing Multiple Choice Science Questions}, 
      author={Johannes Welbl and Nelson F. Liu and Matt Gardner},
      year={2017},
      eprint={1707.06209},
      archivePrefix={arXiv},
      primaryClass={cs.HC},
      url={https://arxiv.org/abs/1707.06209}, 
}

@misc{hendrycks2021measuringmassivemultitasklanguage,
      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},
      year={2021},
      eprint={2009.03300},
      archivePrefix={arXiv},
      primaryClass={cs.CY},
      url={https://arxiv.org/abs/2009.03300}, 
}

@misc{ling2017programinductionrationalegeneration,
      title={Program Induction by Rationale Generation : Learning to Solve and Explain Algebraic Word Problems}, 
      author={Wang Ling and Dani Yogatama and Chris Dyer and Phil Blunsom},
      year={2017},
      eprint={1705.04146},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/1705.04146}, 
}

@misc{pal2022medmcqalargescalemultisubject,
      title={MedMCQA : A Large-scale Multi-Subject Multi-Choice Dataset for Medical domain Question Answering}, 
      author={Ankit Pal and Logesh Kumar Umapathi and Malaikannan Sankarasubbu},
      year={2022},
      eprint={2203.14371},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2203.14371}, 
}

```