gahann/Nepali_law_articles_corpus

Nepal Legal & Government Education Corpus

Dataset Summary

This dataset is a collection of 868 explanatory legal and government-procedure articles scraped from 11 trusted Nepalese sources — legal blogs, law firm publications, government portals, and legal aid / human-rights organizations. It was built as part of NyayaLM, a bilingual (Nepali-English) legal foundation language model for Nepal, where it serves as one of the general-education corpora used to ground the model in citizen-facing legal explanation, government procedure, and administrative guidance — as distinct from primary statutory text (acts, codes, judgments), which is collected separately as NepalLegalCorpus.

This dataset specifically supports the project's Objective 1 (bilingual corpus construction) and downstream Objective 4/5 (instruction-tuning data for legal assistance and conversational capability), by providing the plain-language explanatory material a citizen or law student would actually read, rather than raw statutory language.

Supported Tasks

• Text generation / language modeling — pretraining or continued pretraining on Nepali-English legal-adjacent text.
• Instruction tuning — source material for building question-answer or explanation-style instruction pairs (e.g. "How do I register a company in Nepal?").
• Summarization — article-length legal/government explainer text is suitable for summarization fine-tuning.
• Retrieval-Augmented Generation (RAG) — usable as a grounding corpus for citation-aware legal QA.

Languages

The corpus is bilingual: English (en), Nepali (ne), and code-mixed Nepali-English (ne-en), reflecting how legal and administrative content is actually published in Nepal.

Language	Articles	% of Corpus
en	815	93.9%
ne-en	53	6.1%

Dataset Structure

Data Instances

Each row is one cleaned article:

{
  "title": "How to Register a Private Company in Nepal",
  "text": "Full cleaned article body...",
  "source": "lawalpine.com",
  "url": "https://lawalpine.com/blog/company-registration-nepal",
  "date": "2024-03-11",
  "language": "en",
  "category": "company registration",
  "folder": "Administrative Procedures"
}

Data Fields

Field	Type	Description
title	string	Article title (extracted or inferred from URL)
text	string	Cleaned article body — navigation, ads, sidebars, comments, and share buttons removed
source	string	Source domain the article was scraped from
url	string	Original article URL, for attribution and verification
date	string	Publish date if the source page exposed one, else empty
language	string	en, ne, or ne-en (code-mixed), auto-detected
category	string	Fine-grained legal/administrative topic
folder	string	Top-level corpus grouping used during collection

Data Splits

This release ships as a single train split (868 rows). No evaluation split is defined here — evaluation is handled separately via NepalLegalBench (the project's dedicated benchmark), keeping training and evaluation data sources cleanly separated.

Category Distribution

21 categories, grouped into 6 top-level folders.

Category	Articles	% of Corpus
legal blogs	590	68.0%
company registration	48	5.5%
court procedures	34	3.9%
tax	31	3.6%
divorce	30	3.5%
immigration	25	2.9%
citizenship	18	2.1%
labor law	17	2.0%
marriage registration	14	1.6%
pan/vat	13	1.5%
land registration	12	1.4%
constitutional law	7	0.8%
banking regulations	7	0.8%
human rights	6	0.7%
inheritance	4	0.5%
legal education	3	0.3%
civil law	3	0.3%
government services	3	0.3%
consumer rights	1	0.1%
passport	1	0.1%
local government procedures	1	0.1%

Source Distribution

Top sources by article count (full list of 11 sources included in the dataset metadata):

Source Domain	Articles	% of Corpus
attorneynepal.com	484	55.8%
lawalpine.com	174	20.0%
lawbhandari.com	97	11.2%
lawfirmnepal.com.np	53	6.1%
fwld.org	32	3.7%
neupanelegal.com	16	1.8%
nepallaws.com	5	0.6%
companydartanepal.com	4	0.5%
barristernepal.com	1	0.1%
nitipartners.com	1	0.1%
lawaxion.com	1	0.1%

Corpus Statistics

• Total articles: 868
• Unique source domains: 11
• Categories: 21
• Average article length: ~1588 words
• Date coverage: not consistently available across sources

Dataset Creation

Source Data

Articles were identified by crawling a manually vetted list of trusted domains — government portals, law firm publications, and legal-aid / human-rights NGOs — rather than through generic web search, in order to bias the corpus toward reliable, non-forum, non-AI-generated content, per the project's source-quality requirements.

Collection Methodology

1. Domain vetting — each source domain was manually reviewed before inclusion (government sites, licensed law firms, legal education platforms, and NGOs; forums, social media, and low-quality blogs were explicitly excluded).
2. Crawling — an automated crawler traversed each trusted domain's internal links (breadth-first, robots.txt-respecting) to discover article URLs, rather than relying on a hand-curated link list.
3. Extraction & cleaning — each article's HTML was parsed with trafilatura to strip navigation, advertisements, sidebars, comments, and share widgets, keeping only the substantive article body.
4. Deduplication — article text was hashed and near-identical content (e.g. the same explainer republished across multiple sites) was filtered, keeping a single canonical copy.
5. Language detection — each article was tagged en, ne, or ne-en using script detection (Devanagari presence) combined with automated language identification.
6. Categorization — articles were auto-classified into legal/government topics via keyword matching on URL slug and title, then organized into the folder structure shown above.

Annotations

No manual annotation was performed. Category and language labels are automatically assigned (see Collection Methodology) and have not been human-verified at scale; spot-checking is recommended before high-stakes use.

Personal and Sensitive Information

This corpus consists of published, publicly accessible educational and procedural content. It is not expected to contain personal data about private individuals, but no automated PII scrubbing has been applied since the collection scope (legal-education articles) does not typically involve personal case details. Users working with this data for anything beyond research should independently verify this assumption.

Considerations for Using the Data

Social Impact and Intended Use

This dataset is intended to help ground an AI system in accurate, accessible legal and government-procedure explanation for Nepal, a low-resource language and jurisdiction that is otherwise underserved by existing NLP resources. It is not a substitute for professional legal advice, and content generated from models trained on it should not be treated as authoritative legal counsel.

Known Limitations and Biases

• Source skew: article counts are not evenly distributed across domains (see Source Distribution above); downstream users may wish to rebalance or weight sampling accordingly.
• No legal-accuracy verification: articles reflect the publishing source's interpretation of the law at the time of writing and may be outdated relative to current statutes; this dataset should be paired with primary statutory sources (e.g. NepalLegalCorpus) for anything requiring legal accuracy or citation correctness.
• Automated categorization/language labels are heuristic, not human-verified, and may contain misclassifications.
• Commercial law-firm content may include promotional framing alongside genuine legal explanation.

Licensing Information

Individual articles remain the copyright of their original publishing source; this dataset provides url and source fields for full attribution. This dataset is released for non-commercial research use in connection with the NyayaLM project. Redistribution of the full article text for commercial purposes is not authorized without checking the original source's terms of use. If you are a rights holder and want your content removed, please open an issue on this repository.

Citation Information

If you use this dataset, please cite the NyayaLM project:

@misc{nyayalm2026corpus,
  title        = {Nepal Legal \& Government Education Corpus},
  author       = {NyayaLM Project Contributors},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/gahann/Nepali_law_articles_corpus}},
  note         = {Data collection component of the NyayaLM bilingual legal foundation model project}
}

Dataset Curators

Collected as part of the NyayaLM project (bilingual Nepali-English legal foundation language model for Nepal), data collection component.

Acknowledgements

Thanks to the original publishing sources — government agencies, law firms, legal education platforms, and legal-aid NGOs — whose public educational content made this corpus possible.