bangla-gamba-base

BanglaGamba

BanglaGamba is a custom hybrid language model trained from scratch on a mixed corpus of approximately 9.62B tokens, consisting of Bengali (~7.37B), English (~1.23B), and bilingual translation pairs (~1.03B). The model combines Mamba-3 (State Space Model) and Grouped Query Attention (GQA) to provide strong performance across Bengali language understanding and generation tasks.

Model Details

Resources

• GitHub Repository: ahmed-farhanur-rashid/bangla-gamba
• Training Dataset: datasets/ahmed-farhanur-rashid/bn-foundational-pretrain-corpus

Requirements

This model uses a custom architecture and tokenizer implementation. Loading requires enabling trust_remote_code=True.

Install the required dependencies:

pip install transformers torch mamba-ssm causal-conv1d bnunicodenormalizer

Note: The model was trained on text normalized using bnunicodenormalizer in data pipeline. Running without it may significantly degrade generation quality.

Architecture

Unlike conventional Transformer-only language models, BanglaGamba combines state-space modeling with attention mechanisms:

• Mamba-3 (State Space Model): Efficient linear-time sequential modeling and long-range state recurrence without quadratic attention overhead.
• Grouped Query Attention (GQA): Improves inference efficiency through optimized key-value caching, enhanced with per-head QK-Norm and RoPE.
• SwiGLU FFN: Interleaved in every block for non-linear representation capacity.

This hybrid design aims to balance computational efficiency with strong language modeling performance.

Related Models

BanglaGamba is part of a family of Bengali foundation language models.

• BanglaGSG: tasmin-jahan/bangla-gsg

Limitations

• The model expects text normalized using bnunicodenormalizer, consistent with the preprocessing pipeline used during training.
• Loading requires execution of custom Python modules (modeling_banglagamba.py, configuration_banglagamba.py, tokenization_banglagamba.py) via trust_remote_code=True.
• While primarily trained for Bengali, English support is intended mainly for multilingual understanding and translation-related capabilities.
• As with other large language models, outputs may occasionally be inaccurate or reflect biases present in the training data.

Citation

If you use BanglaGamba in your research, please cite the model:

@misc{banglagamba2026,
  title        = {BanglaGamba},
  author       = {Ahmed Farhanur Rashid},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/ahmed-farhanur-rashid/bangla-gamba}}
}