Bhojpuri Corpus (BhojpuriCorpus) — Monolingual Pretraining Dataset for Bhojpuri (bho)

Overview

BhojpuriCorpus is a monolingual pretraining dataset for the Bhojpuri language (ISO 639-3: bho), containing 386,032 documents and approximately 24.97 Million estimated tokens. It is compiled from multiple public sources and preprocessed for vocabulary training and language model pretraining.

Motivation

BhojpuriCorpus was compiled to aggregate, clean, and deduplicate publicly available Bhojpuri text resources into a single standardized pretraining dataset, facilitating language modeling research for this low-resource Indo-Aryan language.

Dataset Composition

The corpus consists of Devanagari text spanning multiple domains, including general encyclopedic articles, web crawl news and blogs, academic prose, literary writings, and speech transcriptions.

Data Collection

The text content was collected from the following public repositories:

• Wikipedia: Bhojpuri Wikipedia articles.
• Web Crawls: HPLT 2.0 and Goldfish (bho_deva).
• ASR Transcripts: KonthouKabi, Google FLEURS, and Rural Women speech transcripts.
• Parallel Data Targets: Monolingual Bhojpuri subsets of translation datasets (nilayshenai and Google Smol).
• Academic/Literary: JNU-BHLTR corpora, IIT-BHU SSF chunked sentences, and Universal Dependencies Bhojpuri-BHTB treebank.
• Literary Blogs: Domain-specific crawls from GlotCC-V1.

Preprocessing

The aggregated raw text was cleaned using the following steps:

• Instruction Exclusion: Instruction-tuning templates and conversational chat segments were excluded.
• Translation Extraction: For parallel datasets, English translations and alignment metadata were removed, retaining only monolingual Bhojpuri sentences.
• Normalization: Devanagari text was normalized to NFC Unicode. HTML tags, Wikipedia section edit markers, and numeric/punctuation fragments were removed.
• Language Filtering: A grammatical language identification filter was applied to remove Hindi and Awadhi text.
• Deduplication: Normalized text strings were deduplicated globally across all partitions.

Dataset Statistics

• Documents: 386,032
• Words: 18,499,644
• Estimated Tokens (1.35x): 24,974,519
• Vocabulary Size: 344,449
• Language: Bhojpuri (bho)
• Script: Devanagari (Deva)
• File Formats: JSONL, Parquet, CSV

Source Distribution

Source Partition	Document Count	License
fish_food	215,816	CC-BY-4.0
jnu_bhltr	79,884	CC BY-NC-SA 4.0
eng_bho	26,909	CC-BY-4.0
bhojpuri_lr_v3	21,811	CC-BY-4.0
hplt	17,778	CC0
wikipedia	7,603	CC BY-SA 4.0
kumar_bhojpuri	6,567	CC-BY-4.0
iit_bhu	5,254	CC BY-NC-SA 4.0
fleurs_bhojpuri	2,708	CC-BY-4.0
google_smol	844	CC-BY-4.0
glotcc	567	CC-BY-4.0
rural_women	211	CC-BY-4.0
ud_bhtb	83	CC BY-SA 4.0

Dataset Structure

The dataset consists of a single train split. Every document is formatted as a flat dictionary with the following schema:

• id (string): Unique document identifier (bho_corpus_[index]).
• text (string): Preprocessed Bhojpuri text content.
• source (string): Identifier of the original source repository.
• license (string): License governing the source text.
• url (string): Homepage URL of the original source.

JSON Example

{
  "id": "bho_corpus_10250",
  "text": "अउर भाषा सभ के तुलना में भोजपुरी में बहुत कम साहित्य उपलब्ध बा, जवना के मुख्य वजह एकरा के सरकारी मान्यता ना मिलल बा।",
  "source": "wikipedia",
  "license": "CC BY-SA 4.0",
  "url": "https://bh.wikipedia.org"
}

Usage

This dataset can be loaded directly through the Hugging Face Datasets library:

from datasets import load_dataset

dataset = load_dataset(
    "Satyam810/BhojpuriCorpus",
    split="train"
)

print(dataset[0])

Intended Uses

• Tokenizer Training: Training SentencePiece or BPE subword tokenizers.
• Pretraining from Scratch: Pretraining small decoder-only Transformer models.
• Continual Pre-Training: Adapting larger multilingual foundation models to the Bhojpuri language.

Limitations

• Spelling Orthography: Because Bhojpuri lacks a universally standardized orthographic dictionary, there is spelling variation across partitions, particularly between academic collections and spoken transcripts.
• Scale Constraints: At approximately 25 Million estimated tokens, training high-parameter models from scratch is constrained by data size.

Licensing

BhojpuriCorpus is distributed under a mixed open-source license. Individual records are subject to their respective original source licenses (CC0, CC-BY-4.0, CC BY-SA 4.0, or CC BY-NC-SA 4.0), which are documented in the license field of each record. Curation metadata and scripts are licensed under the Apache License 2.0.

Acknowledgements

This dataset aggregates publicly available Bhojpuri resources released by their respective authors and organizations. We gratefully acknowledge the original creators and maintainers whose work made this corpus possible.

Citation

@misc{kumar2026bhojpuricorpus,
  author = {Satyam Kumar},
  title = {BhojpuriCorpus: A Monolingual Pretraining Corpus for Bhojpuri},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/Satyam810/BhojpuriCorpus}}
}