README.md

---
dataset_info:
  features:
  - name: id
    dtype: string
  - name: input
    dtype: string
  - name: opa
    dtype: string
  - name: opb
    dtype: string
  - name: opc
    dtype: string
  - name: opd
    dtype: string
  - name: cop
    dtype: int64
  - name: choice_type
    dtype: string
  - name: exp
    dtype: string
  - name: subject_name
    dtype: string
  - name: topic_name
    dtype: string
  - name: output
    dtype: string
  - name: options
    dtype: string
  - name: letter
    dtype: string
  - name: incorrect_letters
    list: string
  - name: incorrect_answers
    list: string
  - name: single_incorrect_answer
    dtype: string
  - name: system_prompt
    dtype: string
  - name: messages
    list:
    - name: content
      dtype: string
    - name: role
      dtype: string
  splits:
  - name: train
    num_bytes: 221816870
    num_examples: 164539
  - name: test
    num_bytes: 24647517
    num_examples: 18283
  download_size: 144137775
  dataset_size: 246464387
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: test
    path: data/test-*
dataset_name: mkurman/medmcqa-hard
license: cc
language:
  - en
task_categories:
  - multiple-choice
  - question-answering
  - reinforcement-learning
tags:
  - medical
  - MCQ
  - evaluation
  - SFT
  - DPO
  - RL
pretty_name: MedMCQA-Hard
size_categories:
  - 10k<n<1M
---

# medmcqa-hard

**A harder, de-duplicated remix of MedMCQA** designed to reduce memorization and strengthen medical MCQ generalization.

## Why “hard”?

* **Answer list variants:** Each correct option appears in **multiple phrasing/list variants** (e.g., reordered enumerations, equivalent wording), so models can’t rely on surface-form recall and must reason over content.
* **RL-friendly targets:** Every item includes **one canonical correct answer** and both **single** and **set** of incorrect answers → plug-and-play for **DPO**, **RLAIF/GRPO**, and contrastive objectives.
* **Chat formatting:** Adds lightweight **`messages`** (and optional `system_prompt`) not present in the original dataset, making it convenient for instruction-tuned models and SFT.

## Intended uses

* Robust **eval** of medical QA beyond memorization.
* **SFT** with chat-style prompts.
* **DPO / other RL** setups using `single_incorrect_answer` or `incorrect_answers`.

## Data schema (fields)

* `question`: str
* `options`: list[str] (usually 4)
* `letter`: str (A/B/C/D)
* `cop`: int (0-based index of correct option)
* `incorrect_answers`: list[str]
* `single_incorrect_answer`: str
* `messages`: list[{role: "system"|"user"|"assistant", content: str}]
* `system_prompt`: str (optional)

### Example

```json
{
  "question": "Which of the following is true about …?",
  "options": ["A …", "B …", "C …", "D …"],
  "letter": "C",
  "cop": 2,
  "incorrect_answers": ["A …", "B …", "D …"],
  "single_incorrect_answer": "B …",
  "messages": [
    {"role":"system","content":"You are a medical tutor."},
    {"role":"user","content":"Q: Which of the following…?\nA) …\nB) …\nC) …\nD) …"}
  ]
}
```

## Source & attribution

Derived from **MedMCQA** (Pal, Umapathi, Sankarasubbu; CHIL 2022). Please cite the original dataset/paper when using this work.

> **Safety note:** Research/education only. Not for clinical use.