README.md

---
license: cc
language:
- en
tags:
- medical
- spelling
- counting
- qa
- grpo
task_categories:
- question-answering
pretty_name: MedSpellCount-QA
---

# MedSpellCount-QA

## Dataset Summary
**MedSpellCount-QA** is a lightweight dataset for **orthographic counting** framed as question-answering over medical terms.  
Each `input` is a short natural-language question like:

> *“How many **r** are in **warfarine**?”*

The `output` is the **correct count** as an integer (e.g., `1`). This format is convenient for **GRPO** (Group Relative Policy Optimization) or other RL-style post-training, where a simple correctness reward compares multiple candidates per prompt.

*Why a distinct dataset?* Counting letters in real medical vocabulary is a simple, objective task that stresses **spelling attention** and **string reasoning** without requiring external knowledge.

## Use Cases
- **GRPO training**: generate K candidates per prompt and reward exact correctness.  
- **Instruction/QA fine-tuning** for robustness to orthographic queries.  
- **Eval** of character-level attention and tokenization effects on medical terms.

## Languages
- **English** prompts; terms are predominantly **medical**.

## Dataset Structure

### Data Fields
- **input** *(string, required)*: the question, e.g.,  
  `How many 'r' in 'warfarine'?`
- **output** *(integer, required)*: the correct count as text, e.g., `1`.

### Data Instances
```json
{
  "input": "How many 'r' in 'warfarine'?",
  "output": 1
}

```python
from datasets import load_dataset

ds = load_dataset("mkurman/MedSpellCount-QA", split='train')

print(ds)
print(ds[0])
```