Advancing Bengali Next-Word Prediction: A Multi-Domain Text Dataset for Language Modelling and Predictive Text

This is a comprehensive collection of Bengali texts drawn from multiple areas of discourse to support research into Bengali language modeling and next-word prediction. The data was compiled in Summer 2025 from 32 publically available sources which contain a wide variety of writing styles and topics (i.e., national news portals; Wikipedia; Class IX-X textbooks; religious and ethical writings; literature and poetry; blogs; publicly visible social media text; and a small number of clearly labeled examples generated by an artificial intelligence). All data was stored in UTF-8 Bengali Unicode text format and grouped at the source level to enable transparent provenance tracking, domain-based testing, and experimental replication. With respect to size, this combined corpus contained 503,257 tokens and 84,788 unique word forms before preprocessing and 498,023 tokens and 62,009 unique word forms after preprocessing, providing a large body of data for training, benchmarking, and comparative analysis of both statistical and neural methods for Bengali next-word prediction.