metadata
license: apache-2.0
datasets:
- allenai/MADLAD-400
language:
- bn
base_model:
- Qwen/Qwen3-14B-Base
library_name: transformers
Qwen3 14B Base for Bengali: Vocabulary expansion
This model is built on top of Qwen3 14B Base adapted for Bengali using 500M target language tokens sampled from MADLAD-400. It has an additional target vocabulary of 10K.
Model Details
- Vocabulary: This model has an additional target vocabulary of 10K.
- Target vocabulary initialization: The target weights of the embedding and LM head were initialized using mean initialization.
- Training: This model was continually pre-trained on 500M target language tokens sampled from MADLAD-400.
Model Description
- Language: Bengali
- License: Apache 2.0
- Fine-tuned from model: Qwen/Qwen3-14B-Base
Model Sources
- Repository: https://github.com/gucci-j/chat-cve
- Paper: https://arxiv.org/abs/2412.11704
How to Get Started with the Model
Use the code below to get started with the model.
from transformers import AutoTokenizer, AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained(
"atsuki-yamaguchi/Qwen3-14B-Base-bn-madlad-mean-tuned"
)
tokenizer = AutoTokenizer.from_pretrained(
"atsuki-yamaguchi/Qwen3-14B-Base-bn-madlad-mean-tuned"
)
Citation
@article{yamaguchi2025adapting,
title={Adapting Chat Language Models Using Only Target Unlabeled Language Data},
author={Atsuki Yamaguchi and Terufumi Morishita and Aline Villavicencio and Nikolaos Aletras},
journal={Transactions on Machine Learning Research},
issn={2835-8856},
year={2025},
url={https://openreview.net/forum?id=6IdoIKowfe},
note={}
}