|
|
---
|
|
|
license: apache-2.0
|
|
|
language:
|
|
|
- ps
|
|
|
size_categories:
|
|
|
- 10K<n<100K
|
|
|
task_categories:
|
|
|
- text-classification
|
|
|
- text-generation
|
|
|
- question-answering
|
|
|
---
|
|
|
|
|
|
# ZamAI Pashto Dataset (Cleaned)
|
|
|
|
|
|
## Dataset Summary
|
|
|
|
|
|
This repository hosts a cleaned and QC'ed slice of the ZamAI Pashto corpus. The release contains 28,650 Pashto-language articles and prompts that have been deduplicated, normalised, and aligned for instruction-style training as well as prompt/completion workflows.
|
|
|
|
|
|
## Dataset Details
|
|
|
|
|
|
- **Curated by:** ZamAI Team
|
|
|
- **Language(s):** Pashto (ps)
|
|
|
- **License:** Apache-2.0
|
|
|
- **Version:** v1.0 (2025-06-23)
|
|
|
- **Sources:** BBC Pashto, Radio Azadi, community-contributed Pashto corpora
|
|
|
- **Processing Pipeline:** [ZamAI Pashto Data Processing Pipeline](https://github.com/ZamAI-Pashto/ZamAI-Pashto-Data-Processing-Pipeline)
|
|
|
|
|
|
## Dataset Structure
|
|
|
|
|
|
- **CSV splits:** `pashto_cleaned_train.csv`, `pashto_cleaned_val.csv`, `pashto_cleaned_full_dataset.csv`
|
|
|
- **Instruction JSONL:** `pashto_train_instruction.jsonl`, `pashto_val_instruction.jsonl`
|
|
|
- **Prompt/Completion JSONL:** `pashto_train_prompt_completion.jsonl`, `pashto_val_prompt_completion.jsonl`
|
|
|
|
|
|
### Fields
|
|
|
|
|
|
| Field | Description |
|
|
|
| --- | --- |
|
|
|
| `title` | Source headline or generated title |
|
|
|
| `text` | Cleaned Pashto article body |
|
|
|
| `source` | Origin of the example (news outlet / pipeline tag) |
|
|
|
| `prompt` | Instruction-style prompt derived from the article |
|
|
|
| `completion` | Expected model output/completion |
|
|
|
| `instruction` | (JSONL) Instruction text for instruction-tuning |
|
|
|
| `input` | (JSONL) Optional input/context paired with the instruction |
|
|
|
| `output` | (JSONL) Target response |
|
|
|
|
|
|
### Splits
|
|
|
|
|
|
- `train`: 25,785 examples
|
|
|
- `validation`: 2,865 examples
|
|
|
- `full`: 28,650 examples (union of train + validation)
|
|
|
|
|
|
## Accessing the Data
|
|
|
|
|
|
Files are tracked with Git LFS. After cloning, run `git lfs pull` in the repository to download the actual CSV/JSONL payloads.
|
|
|
|
|
|
## Cleaning & Normalisation
|
|
|
|
|
|
1. Dropped rows with empty title/text
|
|
|
2. Removed duplicate content hashes
|
|
|
3. Normalised whitespace and Unicode (NFKC)
|
|
|
4. Filtered samples shorter than 10 characters
|
|
|
5. Generated aligned prompts, completions, and instruction templates
|
|
|
|
|
|
## Intended Uses
|
|
|
|
|
|
- Fine-tuning Pashto T5/mT5 style models
|
|
|
- Instruction-tuning chat assistants for Pashto
|
|
|
- Building evaluation sets for Pashto summarisation and QA
|
|
|
|
|
|
## Limitations
|
|
|
|
|
|
- Dominated by news-domain writing; colloquial data is limited
|
|
|
- Automatically generated prompts/completions may include occasional artefacts—consider manual review before deployment
|
|
|
- Despite cleaning, residual duplicated facts may remain due to mirrored reporting across sources
|
|
|
|
|
|
## Citation
|
|
|
|
|
|
```bibtex
|
|
|
@misc{tasal2025_zamai_pashto_cleaned,
|
|
|
title = {ZamAI Pashto Dataset (Cleaned)},
|
|
|
author = {Yaqoob Tasal and the ZamAI Team},
|
|
|
year = {2025},
|
|
|
howpublished = {\url{https://huggingface.co/datasets/tasal9/ZamAI-Pashto-Dataset-Cleaned}}
|
|
|
}
|
|
|
```
|
|
|
|
