A cleaned Nepali-language text corpus extracted from official PDF documents published by the Ministry of Home Affairs (MoHA), Nepal (moha.gov.np). The corpus covers annual progress reports and quarterly disclosures spanning fiscal years 2076/77 – 2082/83 (approx. 2019–2026).

Each row is a self-contained chunk of Nepali text (~300–1200 characters), cleaned of OCR artifacts and annotated with rich metadata including source URL, content type, and quality metrics.

Dataset Structure

Splits

Split	Rows
train	~1,820
validation	~230
test	~230
total	~2,280

Fields

Field	Type	Description
`text`	string	Cleaned Nepali text chunk (300–1200 chars)
`id`	string	Unique chunk ID — format: `moha-{pdf_id}-{chunk_id}`
`chunk_global_id`	int	Global chunk index across all PDFs
`chunk_local_id`	int	Chunk index within its source PDF
`source_pdf_id`	int	Source PDF number (1–14)
`source_url`	string	Direct URL to the original PDF on moha.gov.np
`source_filename`	string	PDF filename
`source_total_tokens`	int	Total token count of the original PDF (all languages)
`source_nepali_tokens`	int	Nepali-only token count of the original PDF
`char_count`	int	Character count of this chunk
`nepali_char_ratio`	float	Fraction of Devanagari characters in chunk (0.0–1.0)
`content_type`	string	Semantic content category (see below)
`language`	string	ISO 639-1: `ne` (Nepali)
`script`	string	ISO 15924: `Deva` (Devanagari)
`domain`	string	`government`
`country`	string	`NP` (Nepal)
`organization`	string	`Ministry of Home Affairs, Nepal (MoHA)`
`document_type`	string	`annual_progress_report`
`license`	string	`public_domain_gov_np`
`dataset_version`	string	`1.0`
`created_date`	string	Date this dataset was generated (YYYY-MM-DD)

Content Types

Value	Description
`report_narrative`	Main body text of annual reports
`policy_text`	Legal, policy, or regulatory passages
`table_data`	Numeric or tabular content
`table_of_contents`	TOC sections with page references
`abbreviations`	Abbreviation or glossary lists
`appendix`	Annexure / appendix sections
`other`	Miscellaneous short passages

Data Collection & Processing

Source

All documents were downloaded from moha.gov.np — the official website of the Ministry of Home Affairs, Government of Nepal. Documents are publicly available government publications.

PDF Extraction

Text was extracted from PDF files using automated extraction tools. Because the PDFs are primarily scanned government documents, the raw extracted text contained significant OCR noise.

Cleaning Pipeline

The following cleaning steps were applied to every line of text before chunking:

Step	What was removed / fixed
`[Page N]` markers	Extractor-inserted page labels
`(cid:N)` sequences	Corrupt PDF glyph encoding artifacts
TOC dot leaders	Runs of 4+ dots collapsed to `…`
Unicode replacement char `U+FFFD`	Bad bytes from encoding errors
Private-use area chars `U+E000–U+F8FF`	PDF internal glyph references
Box-drawing chars `─`	Separator line artifacts
Cedilla `¸`	OCR noise character
OCR space injection — Rule 1	Space inserted before a vowel sign (matra); matras cannot start a Nepali word, so the space is always spurious. Example: `सशु ासि` → `सशुासि`
OCR space injection — Rule 2	Matra followed by space then a 1–2 character Devanagari fragment — indicates a mid-word split by OCR. Example: `कायहि रू` → `कायहिरू`
Double spaces	Collapsed to single space
NFC normalization	Unicode canonical composition applied

Chunking Strategy

Lines are first grouped into paragraphs (split on blank / very short lines)
Paragraphs are merged until the chunk reaches 300–1200 characters
Chunks below the minimum are appended to the previous chunk
Any chunk where less than 35% of characters are Devanagari is discarded (filters out garbled / non-Nepali content)

Garbled PDF Handling

PDFs #2 and #4 in the source collection had severely corrupted Devanagari encoding ((cid:N) sequences making up >5% of characters). These were automatically detected and skipped — no text from those PDFs appears in the dataset.

Example Row

{
  "text": "गृह मन्त्रालयले आर्थिक वर्ष २०७९/८० मा सार्वजनिक सुरक्षा, सीमा व्यवस्थापन र आन्तरिक सुशासन सम्बन्धी विभिन्न कार्यक्रमहरू सञ्चालन गरेको छ।",
  "id": "moha-001-0042",
  "chunk_global_id": 42,
  "chunk_local_id": 42,
  "source_pdf_id": 1,
  "source_url": "https://www.moha.gov.np/upload/.../annual_report_2079_80.pdf",
  "source_filename": "annual_report_2079_80.pdf",
  "source_total_tokens": 412000,
  "source_nepali_tokens": 5200,
  "char_count": 148,
  "nepali_char_ratio": 0.8716,
  "content_type": "report_narrative",
  "language": "ne",
  "script": "Deva",
  "domain": "government",
  "country": "NP",
  "organization": "Ministry of Home Affairs, Nepal (MoHA)",
  "document_type": "annual_progress_report",
  "license": "public_domain_gov_np",
  "dataset_version": "1.0",
  "created_date": "2025-04-26"
}

Intended Uses

Language model pre-training — Nepali domain-specific corpus
Fine-tuning — government/legal Nepali NLP tasks
Benchmarking — Nepali text generation and understanding
Linguistic research — Devanagari script, Nepali morphology
Low-resource NLP — Nepali is an under-resourced language; this corpus adds government-domain coverage

Limitations

OCR quality varies across PDFs — some text may still contain reconstruction errors that could not be fixed without a Nepali spell-checker
The corpus covers only MoHA documents; it does not represent all domains of Nepali language
Table content (table_data) may have incomplete rows due to PDF column extraction limitations
Some conjunct characters (e.g., कायक्रम vs the OCR form कायक्रि म) may not be fully restored where the base consonants themselves were misread by OCR

License

All source documents are official publications of the Government of Nepal (Ministry of Home Affairs) and are in the public domain under Nepal's government open publication policy. This dataset is released for research and educational use.

Citation

If you use this dataset, please cite it as:

@dataset{moha_nepal_nepali_2025,
  title     = {MoHA Nepal -- Nepali Government Corpus},
  author    = {Ministry of Home Affairs, Nepal (source) and dataset contributors},
  year      = {2025},
  url       = {https://huggingface.co/datasets/YOUR_USERNAME/moha-nepal-nepali},
  note      = {Extracted and cleaned from moha.gov.np official PDF publications}
}

Contact

For questions about this dataset, open a discussion on the dataset page. For questions about the source documents, visit moha.gov.np.

Downloads last month: 21

Total file size:

2.44 MB