metadata
license: apache-2.0
Model Card: bnT5-32k
Model Details
- Model Name: bnT5-32k
- Model Type: Text-to-Text Transformer (T5 architecture)
- Languages: Bangla
- License: Apache-2.0
- Training: Trained from scratch
Model Description
bnT5-32k is a Bangla T5 model trained from scratch using a 32k-vocabulary SentencePiece tokenizer. It is designed for a wide range of Bangla NLP tasks, including summarization and question answering. The model improves token coverage and generation quality compared to multilingual models like mT5.
Acknowledgments
This work was supported by the Google Cloud TPU Research Program, which provided access to TPU resources for large-scale pretraining and experimentation.
Model Sources
- Repository: Github
- Paper: Submitted for publication