Mizo-Language-Corpus-4M

Description

The Mizo-Language-Corpus-4M is an open-source monolingual Mizo dataset containing 4 million sentences, curated by MWireLabs, Meghalaya, India. Derived from a larger 5.94 million-sentence corpus, this dataset has been meticulously processed to support natural language processing (NLP) research, promote linguistic equity, and foster open development in low-resource languages.

Dataset Statistics

• Total Sentences: 4,000,000
• Total Tokens: ~93,000,000
• Average Tokens per Sentence: 23.3
• File Size: ~394MB
• Language: Mizo (Tibeto-Burman)
• Source Corpus Size: 5,944,581 sentences (138.7M tokens)

License

This dataset is released under the CC BY 4.0 license, requiring attribution to MWireLabs and the original sources. Users must cite this dataset and its contributors in any derived work.

Attribution

• Curator: MWireLabs, Meghalaya, India
• Contact: Through HuggingFace
• Sources:
  • IndicNLCorp (WMT-23 IndicNECorp1.0) for a significant portion of the raw text.
  • Hugging Face datasets, including HuggingFaceFW/fineweb.
  • https://dipr.mizoram.gov.in (Directorate of Information & Public Relations, Government of Mizoram, India) for additional press release data.

Citation

Please cite this dataset as:

@misc{mizoCorpus4M2025,
  title={Mizo-Language-Corpus-4M},
  author={MWireLabs},
  year={2025},
  publisher={HuggingFace},
  howpublished={\url{https://huggingface.co/datasets/MWireLabs/mizo-language-corpus-4M}}
}

Additionally, cite the original sources: IndicNLCorp (WMT-23), Hugging Face dataset contributors, and the Directorate of Information & Public Relations, Government of Mizoram.

Processing Details

• Deduplication: Removed exact duplicate lines across the corpus using a set-based approach to ensure uniqueness.
• Cleaning:
  • Filtered noise (e.g., URLs, special characters beyond basic punctuation) with regular expressions.
  • Split long paragraphs into sentences using punctuation boundaries (., !, ?).
  • Removed lines with 3+ repeated characters (e.g., XXX, !!!) and those with 4 or fewer words.
  • Eliminated lines with more than 4 digits to reduce noise.
  • Targeted removal of garbled text (e.g., non-ASCII dense lines) and web artifacts (e.g., Photobucket, blogspot references).
• Validation: Random sampling and manual review confirmed data quality.

How to Use the Dataset

Using the datasets Library

from datasets import load_dataset

# Load the dataset
dataset = load_dataset(
    "MWireLabs/mizo-language-corpus-4M",
    data_files="mizo-language-corpus-4m.txt",
    split="train"
)

# Verify the data
print(f"Total examples: {len(dataset)}")
print(f"\nFirst 3 examples:")
for i in range(min(3, len(dataset))):
    print(dataset[i])
print(f"\nDataset features: {dataset.features}")

Using the Hugging Face Hub CLI

1. Install the Hugging Face CLI:

pip install huggingface_hub

2. Log in:

huggingface-cli login

3. Download the file:

huggingface-cli download MWireLabs/mizo-language-corpus-4M mizo-language-corpus-4m.txt --local-dir ./

Intended Use

This corpus is intended for training and evaluating Mizo language models, including:

• Masked language modeling
• Sentiment analysis
• Named entity recognition
• Text classification
• Machine translation
• Other NLP tasks

It supports academic research, educational tools, and regional language technology development.

Related Resources

• Model trained on this data: Mizo-RoBERTa - RoBERTa-base model for Mizo language
• Organization: MWireLabs on HuggingFace

Limitations

• Monolingual focus may limit bilingual applications.
• Some cultural nuances might be underrepresented due to cleaning filters.
• Dataset represents formal and written Mizo more than conversational language.

Contributions

We welcome feedback, improvements, and additional data contributions. Please contact us through the HuggingFace community discussions or issues section.

---

MWireLabs - Building AI for Northeast India 🚀