BoCorpus

A comprehensive Tibetan corpus dataset for language model training and NLP research.

Dataset Description

BoCorpus is a curated collection of classical Tibetan texts compiled from multiple digital collections. The dataset is designed for training language models and conducting research in Tibetan natural language processing.

Collections Included

The corpus contains texts from the following collections:

• Bon Kangyur: 151 texts
• Derge Kangyur: 103 texts
• Derge Tengyur: 213 texts
• DharmaEbook: 98 texts
• Pagen Project: 1 texts
• Tsadra Collection: 266 texts
• འབྲི་ལུགས་བང་མཛོད་སྐོར་ལྔ།: 136 texts
• རིན་ཆེན་གཏེར་མཛོད་ཆེན་མོ།: 71 texts

Data Statistics

• Total records: 1039
• Total characters: 603,376,880
• Average characters per text: 580,728

Dataset Schema

Column	Type	Description
id	string	Unique UUID4 identifier for each record
collection	string	Name of the source collection
filename	string	Original filename (without extension)
text	string	Full text content with all line breaks removed
char_count	int64	Total number of characters in the text

Usage

Loading with HuggingFace Datasets

from datasets import load_dataset

dataset = load_dataset("openpecha/BoCorpus", split="train")

# Access a single example
example = dataset[0]
print(f"Collection: {example['collection']}")
print(f"Characters: {example['char_count']}")
print(f"Text preview: {example['text'][:100]}...")

Loading with Pandas

import pandas as pd

df = pd.read_parquet("bo_corpus.parquet")
print(df.head())

Loading with PyArrow

import pyarrow.parquet as pq

table = pq.read_table("bo_corpus.parquet")
df = table.to_pandas()

Data Preparation

The texts in this dataset have undergone the following preprocessing:

1. Newline removal: All newline characters (\n) are removed to create continuous text strings
2. UUID assignment: Each text receives a unique UUID4 identifier
3. Character counting: Total character count is computed for each text
4. Collection tagging: Each record is tagged with its source collection name

Citation

If you use this dataset in your research, please cite:

@dataset{bocorpus,
  title = {BoCorpus: A Tibetan Text Corpus},
  author = {OpenPecha},
  year = {2024},
  url = {https://huggingface.co/openpecha/BoCorpus}
}

License

This dataset is released under the MIT License.

Acknowledgments

This corpus was prepared by OpenPecha as part of their mission to make Tibetan Buddhist texts accessible for digital research and AI applications.