Dataset Card for Tagin-mono-v1

Dataset Description

The Tagin-mono-v1 dataset is a compiled, lowercased monolingual corpus of the Tagin language. It is intended primarily for domain adaptation and continued pre-training of language models via Masked Language Modeling (MLM).

Tagin (ISO 639-3: tgj) is a Tibeto-Burman language spoken primarily in the Upper Subansiri district of Arunachal Pradesh, India. It is considered a low-resource and vulnerable language. This dataset is part of an ongoing effort to digitize the language and develop natural language processing (NLP) tools (such as Named Entity Recognition) for the Tagin community.

• Creator: Tungon Dugi
• Institution: National Institute of Technology Arunachal Pradesh
• Email: tungondugi@gmail.com
• Language(s): Tagin (tgj)
• License: cc-by-nc-nd-4.0

Dataset Structure

The corpus contains a total of 115,323 sentences. To guarantee exact reproducibility for benchmark training runs, the data was deterministically combined, lowercased, shuffled (seed 42), and split into train and validation sets.

• Train Split: 109,556 sentences
• Validation Split: 5,767 sentences

Data Sources

The dataset is a comprehensive amalgamation of Tagin text carefully compiled from several raw domains. These sources include community booklets, manual text generation by native speakers, and audio transcriptions. The corpus is designed to capture realistic language use, spanning a variety of contexts such as translated scriptures, educational materials, spoken-word records, and supplementary contributions from the native speaker community.

Data Format

The dataset is provided in jsonl format. Each line is a JSON object with a single text key containing the lowercased Tagin sentence.

{"text": "si-donyi hilo kilin koongv"}
{"text": "atu-aki lo kurium nv kulu bunyi doocho la riumchingjering la riumdo gu doolin tho"}

Language Information: Tagin

• Language Name: Tagin
• ISO 639-3 Code: tgj
• Family: Sino-Tibetan -> Tibeto-Burman -> Tani
• Status: Vulnerable. Tagin is primarily an oral language, and comprehensive digital corpora are extremely scarce. This dataset represents a significant step forward in providing computational and machine learning resources for the language.

Intended Use

This dataset is designed specifically for Domain-Adaptive Continued Pre-training. By continuing to pre-train base multilingual models (such as XLM-RoBERTa, ModernBERT, or IndicBERT) on this corpus using an MLM objective, models learn Tagin syntax, orthography, and vocabulary. This greatly improves their downstream performance on Tagin-specific NLP tasks.

Usage

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("repleeka/Tagin-mono-v1")

# View a sample from the training set
print(dataset["train"][0])
# Output: {'text': '...'}

Preprocessing Details

To ensure uniformity across all text sources for Masked Language Modeling:

• Casing: All text has been strictly lowercased.
• Punctuation: Standard punctuation is retained to allow models to learn sentence boundaries and syntax.
• Loanwords: English and Hindi loanwords commonly used in daily Tagin speech are retained to reflect realistic modern language usage.

Bias, Risks, and Limitations

• Domain Bias: A significant portion of this corpus is derived from translated scriptures, educational community booklets, and curated NER text. As a result, the vocabulary may lean slightly formal or religious compared to everyday colloquial Tagin.
• Dialectal Variation: The dataset primarily captures the dialect spoken in the Upper Subansiri district. Variations spoken in peripheral regions might be underrepresented.

Acknowledgments

We would like to extend our deepest gratitude to the native speakers, elders, and community members who contributed to generating, transcribing, and verifying these texts. This digitization effort represents a collaborative step forward for the Tagin community.

Citation

If you utilize this dataset in your research, please cite it as follows:

@misc{dugi2026taginmono,
  author = {Dugi, Tungon},
  title = {Tagin-mono-v1: A Monolingual Corpus for the Tagin Language},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/repleeka/Tagin-mono-v1}}
}