junaid008/Pashto-100k-Pairs

---
language:
- ps

license: apache-2.0

task_categories:
- text-generation
- question-answering
- summarization
- translation

task_ids:
- dialogue-generation
- dialogue-modeling
- language-modeling
- conversational
- text2text-generation

pretty_name: Qehwa AI Pashto 100K Fine-Tuning Dataset

size_categories:
- 100K<n<1M

tags:
- pashto
- llm
- instruction-tuning
- fine-tuning
- nlp
- conversational-ai
- low-resource-languages
- qehwa-ai
- supervised-finetuning
- dataset

annotations_creators:
- machine-generated
- expert-generated

source_datasets:
- original

language_creators:
- found

multilinguality:
- monolingual
---

# Qehwa AI - Pashto 100K Fine-Tuning Dataset

## Overview

Qehwa AI presents a large-scale Pashto instruction tuning dataset containing **100,000+ high-quality instruction-response pairs** designed for supervised fine-tuning, conversational AI, and downstream NLP tasks.

This dataset was created to advance AI research for the Pashto language, a significantly underrepresented low-resource language spoken by millions worldwide. The dataset covers more than **20 diverse domains** and is curated as part of a broader **1.5 billion token Pashto corpus**.

The primary goal of this dataset is to improve:

- Pashto language understanding
- Instruction following
- Conversational capabilities
- Reasoning and knowledge generation
- Domain adaptation for large language models

---

# Dataset Details

| Field | Value |
|---|---|
| Dataset Name | Qehwa AI Pashto 100K |
| Language | Pashto |
| Dataset Size | 100,000+ instruction-response pairs |
| Domains Covered | 20+ |
| Format | JSONL |
| Task Type | Supervised Fine-Tuning (SFT) |
| Corpus Size | 1.5B+ Pashto Tokens |

---

# Covered Domains

The dataset includes diverse instruction-response examples from domains such as:

- Education
- Health
- Culture
- History
- Technology
- Religion
- Daily Conversations
- Reasoning
- General Knowledge
- Ethics
- Literature
- Translation
- Writing Assistance
- Summarization
- Question Answering
- AI & Programming
- Lifestyle
- Informational Queries

---

# Dataset Structure

Each sample follows a standard instruction tuning format.

## Example

```json
{
"instruction": "د مصنوعي ځيرکتيا تعريف څه دی؟",
"response": "مصنوعي ځيرکتيا هغه ټکنالوژي ده چې کمپيوټرونو ته د انسان په څېر د فکر او زده کړې وړتيا ورکوي."
}
```

---

# Data Collection & Curation

The dataset was carefully curated using multiple stages of generation, filtering, and refinement, including:

- Human-written samples
- AI-assisted generation
- Topic diversification pipelines
- Manual cleaning and filtering
- Deduplication
- Language consistency checks
- Formatting normalization

Special focus was placed on maintaining:

- Natural Pashto phrasing
- Instruction diversity
- Topic balance
- High response quality
- Multi-domain coverage

---

# Intended Uses

This dataset is intended for:

- Fine-tuning Large Language Models (LLMs)
- Instruction tuning
- Conversational AI systems
- Pashto chatbots
- NLP research
- Academic experimentation
- Low-resource language AI development
- Evaluation of multilingual models

---

# Limitations

While extensive efforts were made to ensure quality, the dataset may still contain:

- Occasional factual inaccuracies
- Bias inherited from source data
- Dialectal variations
- AI-generated imperfections
- Uneven topic distributions

The dataset should be reviewed before deployment in production or safety-critical systems.

---

# Ethical Considerations

This dataset was created to support low-resource language communities and improve accessibility of AI technologies for Pashto speakers.

The creators encourage responsible usage and discourage applications involving misinformation, harmful automation, or misuse against vulnerable communities.

---

# Citation

If you use this dataset in your research or projects, please cite:

```bibtex
@dataset{qehwa_pashto_100k,
title={Qehwa AI: Pashto 100K Fine-Tuning Dataset},
author={Junaid Ahmad},
year={2026},
description={A large-scale Pashto instruction tuning dataset containing over 100,000 instruction-response pairs across 20+ domains.},
url={https://huggingface.co/datasets/junaid008/Pashto-100k-Pairs}
}
```

---

# Acknowledgements

Special thanks to the open-source AI community and tools that made this work possible, including:

- Hugging Face
- Unsloth AI
- Open-source LLM ecosystem
- Pashto language contributors and researchers

---

# Future Work

Upcoming goals include:

- Expanding to 1M instruction pairs
- Adding multilingual alignment
- Creating evaluation benchmarks
- Developing Pashto reasoning datasets
- Releasing additional Pashto corpora and models
- Improving dialect coverage

---

# Contact

## Qehwa AI

Advancing Pashto AI and Low-Resource Language Research.