A prompt bank for Swahili-language LLM tasks and evaluation.
# π°πͺ Swahili PromptBank
A curated collection of 200+ high-quality Swahili prompts designed to improve the alignment, evaluation, and training of large language models (LLMs) in low-resource languages.
---
## π About
**Swahili PromptBank** is an open-source dataset of culturally relevant, linguistically accurate prompts written in Swahili. This project aims to enhance representation and model performance in Swahili by providing carefully reviewed prompts across a range of NLP tasks.
Each prompt is crafted to:
- Match natural Swahili tone and context
- Reflect culturally appropriate scenarios
- Be clear, instructive, and useful for alignment training
---
## π¦ Dataset Structure
```
swahili-promptbank/
βββ prompts/
β βββ classification/
β βββ summarization/
β βββ instruction-following/
β βββ conversational/
βββ examples/
β βββ model\_responses\_comparison.md
βββ scripts/
β βββ validation\_utils.py
βββ README.md
βββ CONTRIBUTING.md
````
---
## π§ Prompt Types
| Task Type | Description |
|-----------------------|-------------|
| `instruction-following` | Prompts where the model is guided to complete a task, e.g., "Explain how to cook ugali." |
| `summarization` | Prompts asking for summaries of Swahili text or stories |
| `classification` | Prompts for categorizing emotions, topics, or sentiment in Swahili |
| `conversational` | Roleplay and dialogue-style prompts in a Swahili cultural context |
---
## π§ͺ Sample Prompt
```json
{
"task": "instruction-following",
"language": "Swahili",
"prompt": "Eleza jinsi ya kupika ugali kwa mtu ambaye hajawahi kufanya hivyo.",
"expected_output": "Hatua kwa hatua za kupika ugali, zikielezwa kwa lugha rahisi."
}
````
---
## β
Quality Criteria
All prompts were:
* **Peer-reviewed** for tone and clarity
* Designed with **cultural alignment** in mind
* Written in **native Swahili** without code-switching
* Labeled with clear metadata for task type and usage
---
## π Use Cases β¦