Jamalianpour/persian-wikipedia-instruct

Persian Wikipedia Instruct

A Persian-language instruction-tuning dataset of ~120,000 samples, generated from the Persian Wikipedia (fawiki) article dump. Each article was cleaned to Markdown, chunked by section, and passed to a locally-run Gemma4 model that produced grounded instruction/response pairs across five task types. The result is ready for supervised fine-tuning (SFT) of Persian LLMs.

Heads-up: this is synthetic data. The instructions and answers were written by an LLM from Wikipedia text, then automatically filtered — not human-verified.

Dataset at a glance

Language	Persian (فارسی), fa
Samples	~120,000
Format	JSONL, Alpaca-style (instruction / input / output)
Source text	Persian Wikipedia dump
Generator	Gemma4:26b model via Ollama
Task types	qa, reasoning_qa, summarization, definition, extraction

Usage

from datasets import load_dataset

ds = load_dataset("Jamalianpour/persian-wikipedia-instruct", split="train")
print(ds[0])

Data fields

field	type	description
task_type	string	one of qa, reasoning_qa, summarization, definition, extraction
instruction	string	the task / question, in Persian
input	string	optional context (used for summarization / extraction; empty otherwise)
output	string	the response, in Persian
source_id	int	Wikipedia page id the sample was derived from
source_title	string	Wikipedia article title
source_url	string	link to the source article
section	string	the article section(s) the sample was grounded in

Example

{
  "task_type": "qa",
  "instruction": "دیگو مارادونا که بود؟",
  "input": "",
  "output": "یک فوتبالیست افسانه‌ای اهل آرژانتین که از بزرگ‌ترین بازیکنان تاریخ فوتبال به شمار می‌رود.",
  "source_id": 1234,
  "source_title": "دیگو مارادونا",
  "source_url": "https://fa.wikipedia.org/wiki/دیگو_مارادونا",
  "section": "مقدمه"
}

How it was built

1. Corpus extraction. The fawiki pages-articles dump was streamed and each main-namespace, non-redirect article was converted from wikitext to clean Markdown (headings, lists, and tables preserved; infoboxes, references, categories, and back-matter sections such as «منابع» and «جستارهای وابسته» removed). Stub articles were dropped.
2. Sample generation. Each article was split into section-sized chunks and passed to a Gemma4 model running locally via Ollama, prompted to produce diverse, self-contained instruction samples grounded only in the chunk text.
3. Quality filtering. Generated samples were filtered to remove non-Persian or garbled output, source-leaking phrasing (e.g. «بر اساس متن»), near-empty answers, and duplicate instructions.

Task type distribution

task_type	count
qa	25,037
reasoning_qa	23,199
summarization	23,835
definition	24,102
extraction	23,423

Licensing and attribution

The source text is from Persian Wikipedia, licensed under CC BY-SA 4.0. Because this dataset is derived from that text, it is released under the same license (CC BY-SA 4.0). If you use or build on this dataset:

• Attribute Wikipedia / the Wikimedia Foundation as the source of the underlying text.
• Share alike — derivative datasets must also be CC BY-SA 4.0.

Limitations

• Synthetic and unverified. Samples are model-generated and only automatically filtered. Some answers may contain factual errors or hallucinations not present in the source.
• Inherited bias. Coverage and viewpoint follow Persian Wikipedia, including its topical and demographic gaps.
• Not for factual authority. Suitable for teaching instruction-following and Persian fluency; not a reliable knowledge base on its own.

Citation

If you use this dataset, please cite it and the underlying source:

@misc{persian_wikipedia_instruct,
  title        = {Persian Wikipedia Instruct},
  author       = {Jamalianpour},
  year         = {2026},
  howpublished = {https://huggingface.co/datasets/Jamalianpour/persian-wikipedia-instruct},
}