SiPaKosa-Books: Book-Level Sinhala-Pali Buddhist Corpus

This is the book-level version of the SiPaKosa dataset. Where SiPaKosa-Sent contains individual sentences, this dataset groups all sentences by book so each row represents one complete book's worth of text.

Related dataset (sentence-level): RaniduG/SiPaKosa-Sent

What's in here

File	Description
corpus/full_corpus.csv	All records from SiPaKosa unified into a single flat file
corpus/full_corpus.parquet	Same as above in parquet format
corpus/grouped_corpus.csv	Text grouped and concatenated at book level
corpus/grouped_corpus.parquet	Same as above in parquet format

Dataset Configs

Config	Path	Description
full_corpus	corpus/full_corpus.csv	All sentences from both sinhala and mixed splits
grouped_corpus	corpus/grouped_corpus.csv	One row per book with full text concatenated

Grouping Logic

For source == IFBC, text is grouped by book_name_si (Sinhala book name). For source == Tripitaka, text is grouped by book_name_en (English book name). Sentences within each book are joined with a newline character.

Columns

Column	Description	Example
document_id	Unique UUID per row	3f4a1b2c-...
book_name_si	Sinhala book name (IFBC only)	විශුද්ධිමාර්ගය
book_name_en	English book name (Tripitaka only)	Khuddaka Nikaya
source	Data source	IFBC or Tripitaka
text	Full book text (sentences joined by newline)	මා හට...
language	Language classification	sinhala or pali

Quick Start

from datasets import load_dataset

# Load the grouped (book-level) corpus
ds = load_dataset("RaniduG/SiPaKosa-Books", "grouped_corpus")
print(ds["train"][0])

# Load the flat full corpus
full = load_dataset("RaniduG/SiPaKosa-Books", "full_corpus")
print(full["train"][0])

# Filter Tripitaka books only
import pandas as pd
df = pd.DataFrame(ds["train"])
tripitaka = df[df["source"] == "Tripitaka"]
ifbc = df[df["source"] == "IFBC"]

Relationship to SiPaKosa

The sentence-level dataset RaniduG/SiPaKosa-Sent contains:

• 786,344 total sentences
• Split into sinhala and mixed configs with train/validation/test splits
• Columns include sentence_id, book_category, book_name_si, book_name_en, source, text, language

This dataset (SiPaKosa-Books) is built directly from SiPaKosa by aggregating all text to book level. Use SiPaKosa for sentence-level tasks like classification or language modelling. Use SiPaKosa-Books for document-level tasks like retrieval, summarisation or RAG indexing.

Sources

• IFBC: 16 digitised historical Sinhala Buddhist books
• Tripitaka: 5 Nikayas scraped from tripitaka.online

License

MIT, for research purposes.

Paper

https://arxiv.org/abs/2603.29221