Bangla NLP Corpus · বাংলা ভাষার ডেটাসেট

A curated, deduplicated, license-aware Bangla text corpus, and a GPT-2 model trained from scratch on it.

📚 Dataset 🤖 Model
Loading statistics…

Documents by source

Documents by topic (automatic labels)

Pipeline funnel

Every rejected document is logged with a reason. Nothing is dropped silently.

Splits & subsets

Licensing

All text comes from Wikimedia projects (Wikipedia, Wikisource, Wikibooks, Wikiquote, Wikivoyage) under CC BY-SA 4.0. News websites are not included: their terms don't allow redistribution without permission. Every record keeps its source URL, license and processing version.

Browse, search and filter every split and subset with the Hugging Face Dataset Viewer.

GPT-2 (small) trained from scratch

Loading training log…

Training loss

Validation perplexity (lower is better)

Samples written during training

Data comes from training_log.txt in the model repo, which updates each time a model is published.

This is a small research model (23M parameters) trained only on encyclopedic and literary text. It writes fluent-looking Bangla but often makes up facts. Don't rely on its output.

The model runs in your browser with transformers.js. Nothing is sent to a server. The first run downloads about 34 MB.