Saptak: A Large-scale Multi-Regional Benchmark Dataset for Poly-Dialectal Neural Machine Translation between Standard Bangla and Regional Dialects, and among Regional Dialects

This dataset is a comprehensive parallel corpus developed for poly-dialectal neural machine translation (NMT), aimed at bridging communication between Standard Bangla and its diverse regional dialects. It is the largest multi-dialect Bengali translation dataset to date (April 4th, 2026), covering eight dialects: Standard Bengali, Sylheti, Chittagonian, Barisal, Noakhali, Rangpur, Rajshahi, and Mymensingh. The dataset includes a substantial number of parallel sentences across these dialects, with 11,056 for Standard Bengali, 10,567 for Chittagonian, 6,422 for Sylheti, 5,712 for Mymensingh, 5,000 for Noakhali, 4,270 for Barisal, 891 for Rajshahi, and 655 for Rangpur, significantly expanding available resources for regional Bangla translation, particularly strengthening representation for Chittagonian and Standard Bengali.

==================================================
Table: Comparative distribution of sentence pairs across Standard Bangla and seven regional dialects in existing datasets and the proposed Saptak dataset

Dialect Coverage Comparison Across Bengali NLP Datasets

Ancholik-NER
Std: 3481 | Syl: 3481 | Chit: 3481 | Bar: 3481 | Noa: 3481 | Ran: 0 | Raj: 0 | Mym: 3481

Anubhuti
Std: 2500 | Syl: 2500 | Chit: 2500 | Bar: 0 | Noa: 0 | Ran: 0 | Raj: 0 | Mym: 2500

Dialect BD
Std: 3452 | Syl: 442 | Chit: 577 | Bar: 790 | Noa: 0 | Ran: 655 | Raj: 891 | Mym: 712

Bhasabodh
Std: 980 | Syl: 980 | Chit: 980 | Bar: 0 | Noa: 0 | Ran: 0 | Raj: 0 | Mym: 0

Chatgaiya Alap
Std: 4011 | Syl: 0 | Chit: 4011 | Bar: 0 | Noa: 0 | Ran: 0 | Raj: 0 | Mym: 0

Onubad
Std: 980 | Syl: 980 | Chit: 980 | Bar: 980 | Noa: 0 | Ran: 0 | Raj: 0 | Mym: 0

Vhasantor
Std: 2500 | Syl: 2500 | Chit: 2500 | Bar: 0 | Noa: 2500 | Ran: 0 | Raj: 0 | Mym: 2500

Saptak (Ours)
Std: 11056 | Syl: 6422 | Chit: 10567 | Bar: 4270 | Noa: 5000 | Ran: 655 | Raj: 891 | Mym: 5712