Awesome Bangladesh › Datasets & open data
The data layer of the Bangladeshi open ecosystem: government portals, research corpora on Hugging Face and Kaggle, and international hosts carrying Bangladesh data. Each dataset states its licence where the publisher declares one, and licence-free items are flagged rather than silently listed. Frozen research corpora stay listed because a completed dataset is a finished product, not an abandoned one. When a dataset claims a government source, we check that the source is named and linkable.
Edit this section on GitHub · Last verified 2026-11-10 · Rules: curation policy
24.3 million documents, 147 GB, CC-BY-4.0 — the largest cleanly licensed Bangla text corpus in existence, published by the Dhaka AI company that trained its own LLM family on it. When the corpus behind a model is this open, everyone's Bangla NLP work gets better.
Mozilla's current official release: 1,277 hours of read Bengali speech, CC0, downloadable in one archive. The largest cleanly-licensed Bangla speech corpus available.
465,000 Bangla paraphrase pairs with round-trip translation and quality filtering, published by the BUET NLP group for LREC 2022. Paraphrase data at this scale is what makes Bangla summarisation and semantic-similarity models actually trainable.
A unified benchmark covering 14 tasks — summarisation, QA, NER, sentiment and more — across five Bangla datasets, introduced at ACL 2024 by a Dhaka-based researcher. BanglaBook finally makes it possible to compare Bangla models across the whole NLP spectrum on one leaderboard.
Sentiment analysis on real-world noisy Bangla — transliterated words, spelling errors and code-mixing included — from authors at SUST Bangladesh (LREC 2022). Noisy-text sentiment is the condition Bangla social-media work actually operates in, and this is the reference dataset for it.
Handwritten Bangla recognition data produced under the Enhancement of Bangla Language in ICT (EBLICT) programme of the Bangladesh Computer Council, built by Giga Tech with the Dhaka University Linguistics department — one of the few Bangla NLP datasets with direct Government of Bangladesh backing.
Synthetically generated Bangladeshi license plates in the Unlicense — the plate format is the real Dhaka-registered layout, and synthetic generation means unlimited training data for BD ANPR systems without privacy concerns.
A Dhaka-maintained lexicon-style dataset assigning sentiment polarity to 61,582 Bangla words — useful as ready-made features for classical ML sentiment pipelines and for expanding Bangla stopword/negation handling.
Community-health-question summarisation data for Bangla — real patient-style questions from health forums, summarised by humans. Health-domain Bangla text data is scarce, and this targets the exact language style of Bangladeshi telehealth and helpline traffic.
The bangla-language-processing organisation (GitHub location: Bangladesh) consolidates cleaned Bangla speech recordings for ASR work. Speech corpora are the hardest Bangla data to license cleanly, which makes every consolidation effort worth listing.
From a SUST-trained NLP researcher: labelled Bangla news headlines across topic categories with deep-learning baselines, GPL-3.0. Headline classification is a standard Bangla media-monitoring task and this ships with working reference code.
Built from a 250-million-article raw corpus including Bangladeshi newspapers (Prothom Alo, NTV) and BD Facebook pages, then labelled for hate speech — MIT licensed. Political and religious hate in Bangla online spaces is a documented moderation problem, and this dataset is the most-cited open resource for it.
docs/sections/datasets.md — part of RanaKhan77/Awesome-Bangladesh
Open the guide ↗