MWirelabs/assamese-monolingual-corpus

Assamese Monolingual Corpus (2025)

A high-quality, sentence-level Assamese monolingual dataset containing 1.61 million cleaned, segmented, and deduplicated sentences in Bengali script. This corpus supports Assamese NLP development, language modeling, and public deployment for Northeast India.

---

Dataset Summary

• Language: Assamese (Bengali script)
• Size: 1,613,879 sentences
• Format: Plain text CSV (text column)
• Total tokens: 77,427,585 (using IndicBERTv2 tokenizer)
• License: CC BY-SA 4.0 — attribution required
• Sources:
  • IITB-IndicMonoDoc
  • Samanantar (Assamese side)
  • Assamese poetry and civic texts

---

📊 Token Statistics

• Total tokens: 77,427,585
• Average sentence length: 36.70 tokens
• Median sentence length: 25.00 tokens
• Minimum sentence length: 10 tokens
• Maximum sentence length: 26,987 tokens

Tokenization performed using ai4bharat/IndicBERTv2-MLM-only.

---

Quickstart: Load with 🤗 Datasets

from datasets import load_dataset

ds = load_dataset("MWirelabs/assamese-monolingual-corpus", split="train")
print(ds[0]["text"])

---

Cleaning Pipeline

This corpus was processed using a modular, script-aware pipeline:

1. Unicode normalization
2. Sentence segmentation using Assamese punctuation (।, !, ?, |)
3. Length filtering (≥10 tokens)
4. Boilerplate removal (e.g. news intros, repetitive closings)
5. Script filtering (removal of non-Bengali-script lines)
6. Whitespace normalization
7. Deduplication

---

Filtering Constraints

• Minimum length: Sentences with fewer than 10 tokens were removed.
• Maximum length: No hard limit was applied, but extremely long lines (>200 tokens) were rare and retained for diversity.
• Proportion removed: Length filtering removed approximately 886,146 lines from the raw segmented corpus (~29.5%).

This ensures the corpus favors complete, meaningful sentences while preserving linguistic diversity.

---

Quality Checks & Validation

• Deduplication: Applied exact match filtering after whitespace normalization. Removed ~496,008 duplicate lines.
• Script validation: Filtered lines with <50% Bengali-script characters to remove non-Assamese content.
• Manual sampling: Random samples were manually inspected to confirm removal of boilerplate, non-script lines, and punctuation clutter.
• Final inspection: Token statistics and sentence samples were reviewed to ensure linguistic and structural integrity.

---

Intended Use

This dataset is designed for:

• Assamese language modeling and generation
• Research on low-resource, script-aware language processing

---

Citation

If you use this dataset, please cite it as:

@misc{mwirelabs_assamese_2025,
  title        = {Assamese Monolingual Corpus (2025)},
  author       = {MWire Labs},
  year         = {2025},
  howpublished = {\url{https://huggingface.co/datasets/MWirelabs/assamese-monolingual-corpus}},
  note         = {Cleaned, segmented, and deduplicated Assamese sentences}
}

---

About MWire Labs

MWire Labs builds ethical, region-first AI infrastructure for Northeast India—focusing on low-resource languages and public accessibility.

Learn more at www.mwirelabs.com

---

Contributions & Feedback

We welcome feedback, contributions, and civic collaborations.
Reach out via Hugging Face.