BanglaDINT-T5

A Single T5 model that understands five Bangla regional dialects — in both native Bangla script and romanized Banglish — and converts them to Standard Bangla, English, or identifies the dialect itself. Fine-tuned from BanglaT5 on the published Vashantor benchmark, it beats prior per-dialect specialist baselines on every single dialect, while being one generalist model instead of five separate single-task ones.

Highlights

• Outperforms both published per-dialect baselines (mT5 and BanglaT5, Faria et al. 2023) on normalization BLEU, across all 5 dialects
• One model, 3 tasks: normalize -> Standard Bangla, translate -> English, identify the dialect
• Handles both Bangla script and romanized Banglish input natively
• Fully reproducible: data pipeline, training, and evaluation code are public (see Reproducibility)

🚀 Live Demo

Try this model directly in your browser, no setup required: BanglaDINT-T5 Dialect Studio

Quickstart

from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
from normalizer import normalize  # pip install git+https://github.com/csebuetnlp/normalizer

MODEL_ID = "tamimaungan/BanglaDINT-T5"
model = AutoModelForSeq2SeqLM.from_pretrained(MODEL_ID)
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, use_fast=False)

text = normalize("আসো কোরোহম?")  # Barishal dialect

for task in ["normalize", "translate to english", "identify dialect"]:
    input_ids = tokenizer(f"{task}: {text}", return_tensors="pt").input_ids
    output_ids = model.generate(input_ids, max_length=64, num_beams=4)
    print(task, "->", tokenizer.decode(output_ids[0], skip_special_tokens=True))

Use the same 3 prefixes for Banglish input — just skip the normalize() call, since that pipeline is for Bangla script only.

Benchmarks

Tested on the Vashantor dataset's official held-out test split (1,869 rows, 5 dialects), never used during training or tuning.

Normalization BLEU, vs. published baselines (Bangla script)

Dialect	mT5 (Faria et al.)	BanglaT5 (Faria et al.)	BanglaDINT-T5
Chittagong	36.75	44.03	49.54
Noakhali	37.43	47.38	54.05
Sylhet	51.32	51.08	60.43
Barishal	48.56	53.50	62.81
Mymensingh	64.74	69.06	70.77
Average*	47.76	53.01	59.52

* Average of the 5 published per-dialect scores, computed here — not an overall figure reported by the original paper. mT5 and BanglaT5 above are 5 separate single-task models (one per dialect); BanglaDINT-T5 is one model doing normalization, translation, and dialect ID jointly.

Full results, both scripts

Task	Bangla script	Banglish
Normalize (BLEU / chrF)	59.58 / 79.61	55.21 / 76.68
Translate to English (BLEU / chrF)	41.43 / 61.03	37.72 / 57.56
Identify dialect (accuracy)	89.2%	87.0%

Baseline	Normalize BLEU
Copy input unchanged	10.52
banglat5 zero-shot (no fine-tuning)	0.17

Identify-dialect, per dialect

Dialect	Bangla	Banglish
Barishal	97.6%	96.0%
Mymensingh	97.6%	94.9%
Chittagong	93.9%	90.7%
Noakhali	85.6%	87.2%
Sylhet	71.7%	66.4%

Training

• Base: csebuetnlp/banglat5 (T5, ~247M params), fine-tuned jointly on 6 task variants (3 outputs x 2 input scripts)
• Data: train + validation combined (10,532 rows), with a 10% stratified slice held out for in-run checkpoint monitoring; the official test split was never used for tuning
• Per-example loss averaging (not per-token), so short targets (dialect labels) aren't under-weighted against long targets (full sentences)
• Checkpoint selection by a balanced score (dialect macro-accuracy + normalize chrF + translate chrF), not plain loss
• 8 epochs, bf16, fixed seed (42); evaluated with beam search (beam=4)

Limitations

• Sylhet identification (~70%) is the weakest of the 5 dialects — likely reflects genuine ambiguity in how Sylheti reads once transcribed into standard Bangla script, not purely a model gap
• Recognizes only these 5 dialects; text from other dialects/languages isn't yet explicitly rejected (planned: confidence-based "Unknown" detection as a separate inference-time component, not a trained class)
• Translation BLEU (~38-41) is a solid first baseline, not state-of-the-art

Reproducibility

Dataset pipeline, training script, and evaluation code: github.com/tamimasad/BanglaDialectNet

Dataset & Credits

Trained on the Vashantor dataset (Faria et al.), accessed via its Kaggle mirror. Fine-tuned from BanglaT5 (Bhattacharjee et al., 2023), licensed CC BY-NC-SA 4.0. If you use this model, please cite both:

@misc{faria2023vashantor,
  title         = {Vashantor: A Large-scale Multilingual Benchmark Dataset for Automated Translation of Bangla Regional Dialects to Bangla Language},
  author        = {Fatema Tuj Johora Faria and Mukaffi Bin Moin and Ahmed Al Wase and Mehidi Ahmmed and Md. Rabius Sani and Tashreef Muhammad},
  year          = {2023},
  eprint        = {2311.11142},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL}
}

@inproceedings{bhattacharjee-etal-2023-banglanlg,
  title     = "{B}angla{NLG} and {B}angla{T}5: Benchmarks and Resources for Evaluating Low-Resource Natural Language Generation in {B}angla",
  author    = "Bhattacharjee, Abhik and Hasan, Tahmid and Ahmad, Wasi Uddin and Shahriyar, Rifat",
  booktitle = "Findings of the Association for Computational Linguistics: EACL 2023",
  year      = "2023",
  pages     = "726--735",
  publisher = "Association for Computational Linguistics",
  url       = "https://aclanthology.org/2023.findings-eacl.54/"
}

License

This model is a derivative of BanglaT5 and is released under the same license as required by its ShareAlike terms: CC BY-NC-SA 4.0 — attribution required, non-commercial use only, derivatives must carry the same license.

Author

Tamim Asad Aungan -- GitHub -- LinkedIn