Uyghur Knowledge Base & RAG Dataset

Contains 74 public Simplified Chinese articles from the Salaam Alykum Uyghur topic, packaged as a production-ready Hugging Face dataset with Parquet splits, Markdown article files, retrieval rows, metadata indexes, and an embedding preview layer.

Human Readers / 普通读者： Looking for normal article reading instead of raw data? Open Files and versions -> content and start with content/README.md. Example article: 3337 维吾尔民族身份是原生还是现代建构. For web-style rendering, use GitHub Pages.

Dataset Viewer

• Articles viewer: https://huggingface.co/datasets/qurancn/Uyghur/viewer/articles/subset
• Retrieval viewer: https://huggingface.co/datasets/qurancn/Uyghur/viewer/retrieval/subset
• Embedding preview: https://huggingface.co/datasets/qurancn/Uyghur/viewer/embeddings/subset
• Embeddable viewer URL: https://huggingface.co/datasets/qurancn/Uyghur/embed/viewer/articles/subset

The primary machine-readable files are Parquet shards under data/. The first columns are lightweight identifiers, titles, dates, previews, counts, and URLs; full content_text, content_markdown, and retrieval text are kept later in the schema to keep the Viewer preview usable.

Splits and configs

Config	Split	Rows	Purpose
articles	train	60	Article-level view for training/evaluation experiments with human review.
articles	validation	7	Article-level validation slice.
articles	test	7	Article-level held-out slice.
articles	subset	10	Lightweight preview subset for the Dataset Viewer and demos.
retrieval	train	60	RAG/search rows with one retrieval chunk per article.
retrieval	validation	7	Retrieval validation slice.
retrieval	test	7	Retrieval held-out slice.
retrieval	subset	10	Lightweight retrieval preview subset.
embeddings	subset	74	Public embedding preview layer using deterministic local character n-gram hashing.

Files

• data/articles/*.parquet — article-level rows with title, provenance, readable previews, full text, Markdown, hashes, and source URLs.
• data/retrieval/*.parquet — retrieval/RAG rows optimized for search, citation indexing, and QA experiments.
• data/embeddings/subset-*.parquet — lightweight embedding preview columns: x, y, embedding_32, and source metadata.
• content/*.md — human-readable Markdown article files.
• metadata/article_index.csv and metadata/article_index.jsonl — lightweight index for citation, SQL filtering, and external crawlers.
• metadata/field_schema.json — exact field names and Parquet types.
• metadata/release-v1.0.0.json — release manifest.

Field design for SQL/Data Studio

The dataset is shaped as a queryable product, not only a file dump. Common filter fields include article_id, title, language, date_published, year, primary_topic, topic_tags_text, author, char_count, has_media, media_count, source_url, github_markdown_url, and content_sha256. Use the Hugging Face Dataset Viewer, Data Studio, or SQL Console to filter by year, topic tag, author, media presence, or source URL.

Format truth

This is an article corpus and RAG/retrieval corpus. It is not a native ShareGPT, Alpaca, chat, or instruction-response dataset. It may be converted into instruction-tuning examples later only if a separate conversion pipeline and new dataset card document that transformation truthfully.

Source and provenance

Source topic: https://salaamalykum.com/cn/topic/%E7%BB%B4%E5%90%BE%E5%B0%94

Each row includes source traceability through source_url, canonical_url, date_published, date_modified, author, topic_tags, and content_sha256. The GitHub public archive is https://github.com/salaamalykum/Uyghur and the readable Pages site is https://salaamalykum.github.io/Uyghur/.

Cleaning and generation rules

• Records were extracted only from public Salaam Alykum topic/article pages.
• Article HTML was normalized into Markdown and plain text.
• Media URLs are kept as public links; media binaries are not republished in this dataset.
• Raw SQL backups, user tables, passwords, tokens, cookies, sessions, IP logs, private tickets, and server secrets are excluded.
• Content hashes are SHA-256 values for integrity and deduplication.
• Parquet shards were written with small row groups and page indexes for stable preview/search behavior.

License and rights

license: other. Article bodies remain source-specific because many posts are translated or reposted public internet material. Do not assume commercial training rights for all article bodies. Generated metadata, schemas, indexes, and packaging files may be reused under the public repository notice where the maintainer has rights to license them.

Recommended uses

• Public-interest search and citation indexing.
• RAG retrieval experiments with human review.
• Media monitoring and public-web archival research.
• Chinese-language Islamic, Uyghur, East Turkistan/Xinjiang, and Muslim human-rights topic exploration.

Limitations and failure boundaries

• Small corpus: 74 articles and 74 retrieval chunks.
• Primary language: Simplified Chinese (zh-Hans).
• Some source posts are reposted or translated public internet material; always cite original source URLs where possible.
• This dataset is not a legal claim, not a private database, and not a comprehensive representation of all Uyghur-related material.
• Empty or very short source articles are preserved when they existed publicly, with hashes and metadata intact.

Versioning

Current release: v1.0.0 for source batch v2026.07.30-uyghur-topic-74. Cite a version tag or commit revision for reproducible downstream work.

Citation

@dataset{qurancn_uyghur_2026,
  author    = {Salaamalykum},
  title     = {Uyghur Knowledge Base & RAG Dataset},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/qurancn/Uyghur},
  version   = {v1.0.0}
}