BANGLA-BOOK

Dataset Summary

BANGLA-BOOK is a collection of Bengali-language educational PDF documents, compiled to support the development of Bengali NLP and language models. The corpus spans multiple academic levels and disciplines — including secondary (SSC), higher secondary (HSC), and tertiary/professional programs (BBA, MBA, PGDM, B.Ed, M.Ed, and BA in Social Science) — sourced from publicly accessible open educational material.

The primary goal of this dataset is to provide a large, structured body of native Bengali text for tasks such as language model fine-tuning, continued pretraining, OCR benchmarking, and general Bengali NLP research, in a language that remains significantly under-resourced relative to its number of speakers.

Dataset Structure

All files are stored as PDFs in a single flat directory:

data/
├── <filename>.pdf
├── <filename>.pdf
├── <filename>_<uuid5>.pdf   # collision-renamed duplicate filename
└── ...

Where two source files shared an identical filename across different program folders, the later file was renamed with a random 5-character suffix to avoid overwriting (e.g. Unit-01_a3f9c.pdf).

Data Fields

Each entry is a single PDF file containing:

• Bengali-language textbook or coursebook content
• Typically organized into chapters/units matching the original curriculum structure
• Mixed content types: narrative prose, expository/textbook prose, and some technical/quantitative material (e.g. economics, statistics)

Data Splits

This dataset is not split into train/validation/test — it is provided as a raw document corpus. Downstream users are expected to perform their own text extraction, cleaning, chunking, and splitting appropriate to their task.

Source Data

Source

Documents were collected from publicly browsable open educational resource directories covering multiple academic programs and levels. No paywalled, access-restricted, or private content was included — all source files were retrieved from open, publicly listed directories.

Collection Process

Files were retrieved via automated crawling of public directory listings, preserving original filenames where possible. No modification was made to the content of the PDFs themselves; only file organization (flattening into a single directory) and, where needed, collision-safe renaming were applied.

Intended Uses

This dataset is intended for:

• Bengali language model pretraining and fine-tuning research
• Bengali OCR and document-layout benchmarking
• Bengali NLP tooling development (tokenization, embeddings, retrieval)
• Academic and non-commercial research into low-resource language modeling

Licensing and Usage Notes

Important: The source documents are educational materials originally published for public educational access. This dataset repository redistributes them for research and educational purposes. Copyright in the underlying textbook content remains with the original publishing institution/authors; this dataset does not grant additional rights beyond those of the original public source.

Users are responsible for reviewing applicable copyright and fair-use/fair-dealing considerations in their jurisdiction before using this data for purposes beyond research, and especially before any commercial use or redistribution of the extracted text itself.

Limitations and Biases

• Domain skew: content is dominated by formal, curriculum-driven textbook prose, which does not represent the full range of contemporary spoken or informal Bengali.
• OCR dependency: any text extracted from these PDFs (rather than used as raw documents) may contain errors depending on the OCR/extraction method used downstream, particularly for documents with complex conjunct characters or degraded scan quality.
• Coverage: while broad across program types, this is not a comprehensive corpus of Bengali educational material — content depends on what was publicly listed by the source at the time of collection.

Citation

If you use this dataset, please cite this repository:

@misc{banglabook,
  title  = {BANGLA-BOOK: A Bengali Educational PDF Corpus},
  author = {LAYEK-143},
  year   = {2026},
  howpublished = {\url{https://huggingface.co/datasets/LAYEK-143/BANGLA-BOOK}}
}

Contact

For questions, corrections, or takedown requests regarding specific source material, please open a discussion on this repository's Community tab.