README.md

---
dataset_info:
  features:
    - name: id
      dtype: string
    - name: question
      dtype: string
    - name: answer
      dtype: string
    - name: language_country
      dtype: string
  splits:
    - name: train
    - name: validation
    - name: test_holdout
  download_size: 25400000
  dataset_size: 29500000
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: validation
        path: data/validation-*
      - split: test_holdout
        path: data/test_holdout-*
license: cc-by-4.0
task_categories:
  - question-answering
  - conversational
language:
  - ak
  - am
  - lg
  - sw
  - en
tags:
  - health-nlp
  - medical
  - adolescent-health
  - sub-saharan-africa
  - multilingual
  - african-languages
pretty_name: AfriHealth-QA
---

# AfriHealth-QA

**A Multilingual Health Question-Answering Dataset for African Languages**

## Dataset Description

This dataset provides a curated collection of health question-and-answer pairs focused on **Adolescent Sexual and Reproductive Health (ASRH)**, maternal care, and infectious disease dynamics across Sub-Saharan Africa. Curated via crowdsourcing vectors by the HASH Consortium, it bridges linguistic variations across local vernacular, code-mixed environments, and standard formats.

### Domain Coverage

*   **Languages**: Akan (`Aka_Gha`), Amharic (`Amh_Eth`), Luganda (`Lug_Uga`), Swahili (`Swa_Ken`), and contextual English variants (`Eng_Uga`, `Eng_Gha`, `Eng_Eth`, `Eng_Ken`).
*   **Countries**: Ghana, Ethiopia, Uganda, Kenya
*   **Core Features**: Real-world colloquial phrasings, dense morphological code-switching, and clinical consensus answers.

## Dataset Summary

| Split | Examples | Description |
|-------|----------|-------------|
| Train | ~29,815 | Labeled question-answer pairs |
| Validation | ~6,686 | Labeled question-answer pairs |
| Test (Holdout) | ~2,618 | Unlabeled questions for private benchmarking |
| **Total** | **~39,119** | |

## Data Schema

| Column | Type | Description |
|--------|------|-------------|
| `id` | string | Unique identifier (format: `ID_XX_Language_Country_Hash`) |
| `question` | string | Health-related question in the specified language |
| `answer` | string | Expert answer (available in train/val, not in test) |
| `language_country` | string | Language-country code (e.g., `Aka_Gha` = Akan, Ghana) |

## Language Reference

| Code | Language | Country | ISO 639 |
|------|----------|---------|---------|
| `Aka_Gha` | Akan (Twi) | Ghana | `ak` |
| `Amh_Eth` | Amharic | Ethiopia | `am` |
| `Eng_Eth` | English | Ethiopia | `en` |
| `Eng_Gha` | English | Ghana | `en` |
| `Eng_Ken` | English | Kenya | `en` |
| `Eng_Uga` | English | Uganda | `en` |
| `Lug_Uga` | Luganda | Uganda | `lg` |
| `Swa_Ken` | Swahili | Kenya | `sw` |

## Data Source

This dataset originates from the **[Zindi Africa](https://zindi.africa)** Multilingual Health QA Challenge, curated for training multilingual health chatbots serving African communities.

## Intended Use

- **Training** multilingual health question-answering models
- **Fine-tuning** LLMs for African health communication
- **Benchmarking** NLP models on multilingual health QA
- **Research** in low-resource language NLP and health informatics

## Key Features

- **8 language-country subsets** covering Akan, Amharic, Luganda, Swahili, and English
- **Code-mixed text**: Natural switching between local languages and English
- **Expert answers**: Long-form, detailed responses (avg ~76 words)
- **Topic-focused**: All questions relate to adolescent sexual and reproductive health

## Example

```
Language: Akan (Ghana)
Question: Ɔkwan bɛn so na mmabunbɛtumi aboa wɔn mfɛfoɔ a nsa anaa nnubɔne ama wɔayɛ wɔn ayayadeɛ?
Answer: Mmabun betumi aboa atipɛnfo a ebia nsa anaa nnubɔne ama wɔayɛ wɔn ayayadeɛ so denam: Nkate fam mmoa a wɔde bɛma...
```

## Usage

```python
from datasets import load_dataset

dataset = load_dataset("ImhotepSystems/AfriHealth-QA")

# Access splits
train_data = dataset["train"]
val_data = dataset["validation"]
test_data = dataset["test_holdout"]

# Filter by language
akan_data = dataset["train"].filter(lambda x: x["language_country"] == "Aka_Gha")

# Example usage
for example in dataset["train"].select(range(3)):
    print(f"Q: {example['question']}")
    print(f"A: {example['answer'][:100]}...")
    print()
```

## Training Models

This dataset is designed to train **imhotep-healthqa** models — open-source health QA systems for African languages:

```python
# Fine-tuning example (pseudocode)
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("ImhotepSystems/imhotep-healthqa")
tokenizer = AutoTokenizer.from_pretrained("ImhotepSystems/imhotep-healthqa")

# Fine-tune on AfriHealth-QA
# ...
```

## Citation

```bibtex
@misc{afrihealthqa2026,
  title={AfriHealth-QA: A Multilingual Health Question-Answering Dataset for African Languages},
  author={Imhotep Systems},
  year={2026},
  publisher={HuggingFace},
  url={https://huggingface.co/datasets/ImhotepSystems/AfriHealth-QA}
}
```

## License

Please refer to the original dataset terms from Zindi Africa. If you use this dataset, please cite both the original source and Imhotep Systems.

## Contact

**Imhotep Systems** — Building open-source AI for African health.
- HuggingFace: [ImhotepSystems](https://huggingface.co/ImhotepSystems)