README.md

---
license: cc-by-nd-4.0
---

# Dataset Card for ZamAI Pashto Processed Dataset

## Dataset Summary

The ZamAI Pashto Processed Dataset provides 28,650 carefully curated Pashto-language records that were collected, cleaned, and normalized through the ZamAI Pashto Data Processing Pipeline. It enables reproducible experimentation for Pashto NLP tasks spanning instruction tuning, summarization, and general sequence-to-sequence modelling.

## Dataset Details

- **Curated by:** ZamAI Team
- **Language(s):** Pashto (ps)
- **License:** CC-BY-ND-4.0
- **Version:** v1.0
- **Last updated:** 2025-06-23
- **Source(s):** BBC Pashto, Azadi Radio, public Pashto corpora, community submissions
- **Pipeline Source:** [ZamAI Pashto Data Processing Pipeline](https://github.com/ZamAI-Pashto/ZamAI-Pashto-Data-Processing-Pipeline)

## Dataset Structure

- **Formats:**
  - CSV: `pashto_cleaned_full_dataset.csv`, `pashto_cleaned_train.csv`, `pashto_cleaned_val.csv`
  - Instruction-tuning JSONL: `pashto_train_instruction.jsonl`, `pashto_val_instruction.jsonl`
  - Prompt-completion JSONL: `pashto_train_prompt_completion.jsonl`, `pashto_val_prompt_completion.jsonl`
- **Fields:**
  - CSV: `title`, `text`, `source`, `prompt`, `completion`
  - Instruction JSONL: `instruction`, `input`, `output`
  - Prompt-completion JSONL: `prompt`, `completion`
- **Splits:**
  - `train`: 25,785 samples
  - `validation`: 2,865 samples
  - `full`: 28,650 samples (CSV + JSONL variants share the same counts)

## Accessing the Data

Large files are stored with Git LFS. After cloning, run `git lfs pull` inside the repository to materialise the CSV and JSONL payloads. Without this step you will only see lightweight pointer files.

## Data Collection Process

- **Gathering:** Pashto language data was automatically collected from diverse online sources, including news websites (e.g., BBC-Pashto), public corpora, and open-access Pashto text repositories. The pipeline utilizes custom Python scripts to crawl, download, and aggregate raw textual data relevant for natural language processing tasks.

- **Cleaning:** The cleaning process removes duplicate entries, irrelevant text, corrupted files, and non-Pashto content. Additional steps include eliminating extra whitespace, fixing encoding issues, stripping HTML tags or special symbols, and filtering out samples below a minimum length threshold to ensure quality and consistency.

- **Normalization:** The text is standardized using Unicode normalization (NFKC), consistent sentence segmentation, and uniform punctuation. Pashto-specific characters and diacritics are normalized, and whitespace is harmonized across samples. The pipeline also optionally standardizes casing and applies consistent formatting to prepare the data for downstream tasks.

- **Tools Used:** Python, pandas, regular expressions (`re`), and custom data processing scripts contained within the [ZamAI-Pashto-Data-Processing-Pipeline](https://github.com/ZamAI-Pashto/ZamAI-Pashto-Data-Processing-Pipeline). Jupyter Notebooks are used for exploration, prototyping, and quality assurance.

## Intended Use

- Fine-tuning Pashto seq2seq and causal language models
- Training instruction-following Pashto assistants
- Building evaluation sets for translation, summarisation, and dialogue experiments

## Limitations and Considerations

- Coverage is skewed toward news-style prose; conversational utterances remain limited.
- Automated cleaning can occasionally trim salutations or remove markup remnants—manual spot checks are encouraged for high-stakes use.
- PIIs are filtered heuristically. Downstream deployments should still review outputs for sensitive details.

## Citation

If you use this dataset, please cite:

```bibtex
@misc{zamai_pashto_processed_2025,
  title = {ZamAI Pashto Processed Dataset},
  author = {ZamAI Team},
  year = {2025},
  howpublished = {\url{https://huggingface.co/datasets/tasal9/ZamAI_Pashto_Dataset}}
}