---
dataset_info:
  features:
  - name: id
    dtype: string
  - name: dataset
    dtype: string
  - name: question
    dtype: string
  - name: options
    sequence: string
  - name: answer
    dtype: string
  splits:
  - name: train
    num_bytes: 2268293
    num_examples: 10687
  download_size: 1254741
  dataset_size: 2268293
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
---
This dataset was constructed as part of the EPFL Modern NLP (MNLP) course project to train and evaluate large language models on **multiple-choice question answering (MCQA)** tasks focused on scientific reasoning.

It aggregates and reformats **10,687 unique examples** from five high-quality academic and biomedical QA datasets, applying consistent structure, question normalization, and cross-source deduplication.

 ### 📊 Dataset Composition

  | Source Dataset | Link | Questions Used | Description |
  |----------------|------|----------------|-------------|
  | ARC-Challenge  | [ai2_arc](https://huggingface.co/datasets/ai2_arc) | 1,119 | Harder science exam questions requiring multi-step reasoning |
  | ARC-Easy       | [ai2_arc](https://huggingface.co/datasets/ai2_arc) | 2,251 | Simpler science questions at the elementary/middle school level |
  | QASC           | [qasc](https://huggingface.co/datasets/qasc)       | 3,000 (subset) | A filtered and deduplicated subset of the QASC dataset, which was originally larger (~8,000+ examples). Only 3,000 unique and diverse questions were selected for balance |
  | OpenBookQA     | [openbookqa](https://huggingface.co/datasets/openbookqa) | 3,317 | 4-option science questions, filtered to keep `humanScore ≥ 1` |
  | PubMedQA       | [pubmed_qa](https://huggingface.co/datasets/pubmed_qa) | 1,000 | Biomedical questions with Yes/No/Maybe answers based on PubMed abstracts |

  ### 🧪 Preprocessing Pipeline

  - **Normalization**: All questions were lowercased and stripped of whitespace for consistency.
  - **Deduplication**: Each question was hashed (`md5(lowercase question)`) to detect and eliminate duplicates across datasets.
  - **Filtering**:
    - OpenBookQA was filtered to retain only questions with `humanScore ≥ 1`.
    - PubMedQA was filtered to retain only labeled questions with answers in {yes, no, maybe}.
    - QASC was **sampled and capped** at 3,000 unique questions to ensure dataset balance.
  - **Unified formatting**: All entries follow the same JSON schema across sources.

  ### 📦 Format

  Each sample follows this structure:

  ```json
  {
  "id": "qasc_481",
  "dataset": "qasc",
  "question": "What do bees use to make honey?",
  "options": ["nectar", "pollen", "water", "leaves"],
  "answer": "A"
}