d3b4g/dhivehi-instruct

Dhivehi Instruct v1

A high-quality, multi-task instruction dataset designed to improve Natural Language Processing (NLP) capabilities for the Dhivehi language. Formatted in a standard conversational structure, it provides clean, contextually accurate data for training, fine-tuning, and evaluating language models on Dhivehi-specific tasks.

v1 contains no synthetic text. Every assistant output is human-written corpus text or a deterministic, rule-based transform of it. No LLM wrote a single Dhivehi character in this dataset — a deliberate design choice for a first release, so the data carries no model-generated hallucination or style contamination.

• Language(s): Dhivehi (Thaana script), English, Romanized Dhivehi (Latin script) — used in the instructions; assistant outputs are Dhivehi or its Latin transliteration.
• Format: Instruction-following (USER / ASSISTANT)
• Size: 20,907 rows — 20,489 train / 418 validation

Dataset Structure

Tasks Included

Task	Rows	Description
headline	8,000	Summarizing news articles into concise, culturally and contextually accurate journalistic headlines. Output is the article's original human-written title.
transliteration	6,000	Converting native Thaana script into Romanized Dhivehi (Latin script), mapping phonetic nuances, suffixes, and double vowels via a rule-based romanizer.
classification	3,915	Categorizing Dhivehi text into standard topics. Prompts include open-ended labeling and multiple-choice constraints, delivered in Dhivehi, English, and romanized Dhivehi.
continuation	3,000	Autoregressive completion that naturally continues a given Dhivehi narrative or news report, using the document's actual next sentences.

Classification is fully balanced: 9 classes × 435 rows — ޚަބަރު (news), ކުޅިވަރު (sports), ވިޔަފާރި (business), މުނިފޫހިފިލުވުން (entertainment), ދިރިއުޅުން (lifestyle), މަޢުލޫމާތު (reference), ސަރުކާރު (government), ދީން (religion), އަދަބިއްޔާތު (literature). About 40% of classification rows are phrased as multiple-choice.

Instruction wording is sampled from hand-authored templates across three languages: Dhivehi (70%), English (20%), and romanized Dhivehi (~10%).

Data Instances

An example of a transliteration task from the dataset:

USER: ތިރީގައިވާ ދިވެހި ޖުމްލަ ތާނައިން ލެޓިނަށް ބަދަލުކޮށްދީ.
ރަށްފަޅުހުރާ ފަޅުގައި ހިންގާ ޕަޓީނާ މޯލްޑިވްސް ރިސޯޓުގައި ނުރައްކާތެރި އެކްސިޑެންޓެއް ހިނގައި އެތަނުގެ މުވައްޒަފަކު އިއްޔެ މަރުވެއްޖެ އެވެ.
ASSISTANT: raffalhuhuraa falhugai hingaa pateenaa moaldivs risoatugai nurakkaatheri eksidente' hingai ethanuge muvazzafaku iyye maruvejje eve.

Row Schema

{
  "messages": [
    {"role": "user", "content": "<instruction + input text>"},
    {"role": "assistant", "content": "<human-written or rule-derived output>"}
  ],
  "task": "headline | classification | continuation | transliteration",
  "instr_lang": "dv | en | latin",
  "template_id": "<instruction template used>",
  "source_doc_id": "<id of the source document>",
  "generator": "template",
  "meta": {}
}

Sources

• d3b4g/dhivehi-corpus — Dhivehi news corpus (~430k docs)
• d3b4g/dhivehi-stories — Dhivehi creative writing (~18k docs)

Construction & Data Hygiene

The build is fully seeded and reproducible. Key steps:

1. Filter — documents need ≥85% Thaana among their letters and 80–3,000 words. Excerpt-length documents (15–79 words) are admitted for the classification task only.
2. Frozen holdout — 5% of all filtered document IDs (22,248) are frozen before any rows are built. No training or validation row derives from a holdout document, and the publishing script hard-fails if any does. The holdout seeds a future DhivehiEval benchmark, giving this dataset a hard-enforced train/eval separation.
3. One task per document — each source document contributes to at most one task, preventing near-duplicate leakage (e.g. a headline row and a continuation row derived from the same article).
4. Validation & dedup — schema, length, and Thaana-ratio checks, followed by exact and MinHash near-duplicate filtering on outputs.

Task-specific cleaning applied during validation:

• Cut-off removal: Some source texts are scraped teasers that end mid-thought with an ellipsis (…). Rather than repair them, rows whose input ends this way are dropped entirely — the published data contains no ellipsis-truncated prompts.
• Script normalization: Arabic ligatures embedded in Dhivehi text (e.g. ﷲ) are romanized in transliteration targets to their standard Latin forms (allah, or abdhullah when attached to a name stem), and Arabic punctuation is mapped to ASCII. Any sentence with a character the romanizer cannot fully convert is dropped, so no stray non-Latin glyphs survive in transliteration outputs.
• Clean outputs: ASSISTANT responses contain strictly the requested output, with no conversational filler — optimal for programmatic evaluation and automated pipelines.

Use Cases

A foundational resource for downstream Dhivehi NLP, including:

• Instruction tuning — teaching base models to follow Dhivehi instructions across classification, generation, and transliteration.
• Content categorization — tagging local-language news, feedback, and unstructured business text using the balanced classification data.
• Accessibility & tooling — leveraging the Thaana↔Latin transliteration pairs for text-to-speech front-ends, search, and reading assistants.

Known Limitations

• Continuation flow: Because outputs are the source document's real subsequent sentences, models trained on this subset may occasionally see abrupt topic transitions where the original article changed context. Evaluate generative output flow accordingly.
• Transliteration convention: Outputs follow one common informal Malé Latin convention; Dhivehi romanization has no single standard, and other schemes exist.
• Classification labels derive from publisher section metadata, which is noisy at the margins.
• Register skew: the corpus is heavily news-weighted; the stories source provides the only substantial non-journalistic register.
• No open-ended QA, summarization, or reasoning tasks yet — planned for v1.1 after a human-reviewed pilot.

License and Intended Use

Source texts are the copyright of their original publishers; this dataset is released for research use in Dhivehi NLP. If you are a publisher of source material and have concerns, please open a discussion on this repository.