paragraphs

A multi-register, long-context text corpus for the Hmar language (hmr, ISO 639-3), part of the Zo languages family.

It contains 32,022 training sequences and 1,000 evaluation sequences (~3.86 million words), formatted into complete paragraphs up to 512 tokens.

Maintained by the Hmar Heritage Foundation (hmarheritage.pages.dev).

---

Overview

• Language: Hmar (hmr, ISO 639-3, Glottolog: hmar1241)
• Family: Zo Languages
• Volume: 32,022 train | 1,000 test (33,022 total sequences, ~3.86M words)
• Sequence Length: Up to 512 tokens (average ~215 words for literary/scripture, ~36 words for comments)
• Segmentation: Complete sentences with natural paragraph boundaries; zero mid-sentence cutoffs
• Validation: Filtered with the hmaraniam language detector (confidence ≥ 0.60)
• Format: Parquet & JSONL
• License: MIT

---

Register Distribution

Linguistic Register	Sequences	% of Sequences	Words	% of Words	Description
Classical & Canonical Scripture	7,962	24.1%	1,689,829	43.8%	Complete chapters packed across classical and modern scripture translations
Literary Prose & Essays	5,878	17.8%	1,291,032	33.5%	Cultural essays, historical chronicles, and literary op-eds
Cultural Reflections & Sacred Verse	1,061	3.2%	228,784	5.9%	Traditional sacred hymns, poetic reflections, and cultural essays
Contemporary Community Commentary	18,121	54.9%	645,518	16.7%	Intact multi-sentence community commentary (online cultural channels ≥ 20 words)
Total	33,022	100.0%	3,855,163	100.0%	~3.86 Million Words

---

Dataset Structure

Following the Foundation's standardized schema, each row is a simple text field containing a coherent paragraph:

{"text": "Chu zan chun lung an lêng deu rieu leiin a ruolnu chu hla rikawt insaktir dingin kan ngên a..."}

---

Usage

from datasets import load_dataset

# Load from Hugging Face
dataset = load_dataset("hmar-heritage-org/paragraphs")

print(f"Train sequences: {len(dataset['train']):,}")
print(f"Test sequences:  {len(dataset['test']):,}")
print("Sample text:", dataset['train'][0]['text'][:120])

---

Source & Provenance

Curated, normalized, deduplicated, and balanced across classical scripture translations, cultural essays, historical prose, sacred verse, and modern community discourse.

Language identification and dialect purity verification are enforced using the Foundation's hmaraniam engine, with full Unicode NFC normalization and standardized punctuation boundaries.

---

Limitations & Sociolinguistic Overlap

While all sequences undergo strict filtering via the Foundation's hmaraniam detector:

• Cross-Dialectal Overlap: Due to the close phylogenetic kinship and mutual intelligibility between Hmar (hmr) and sibling Zo varieties—particularly Mizo (lus)—passages lacking dialect-exclusive markers can overlap across varieties, or feature colloquial particle shifts.
• Digital Code-Switching: Community discussions from web sources naturally reflect modern bilingualism and code-switching across Northeast regional languages.
• Dialect Disambiguation: Passages containing distinctive morphology and vocabulary leverage hmaraniam's frequency-weighted exclusive particle and lexicon tables to enforce high Hmar dialectal purity.

---

About & Stewardship

Curated and maintained by the Hmar Heritage Foundation (hmar-heritage-org) as an open, sovereign linguistic resource for the Hmar language and the broader Zo languages family.