Gemma 3 PT 4B Continually Pretrained on Tibetan
This model is a continual pretraining (CPT) checkpoint built by further pretraining Gemma 3 PT 4B on the Tibetan portion of the MC^2 Corpus.
The model is intended to improve Tibetan language modeling and to support research on low-resource language adaptation.
Training details and methodology are described in: "Efficient Low-Resource Language Adaptation via Multi-Source Dynamic Logit Fusion" (ACL 2026).
Training Data
- Corpus: Tibetan subset of MC^2 Corpus
- Language: Tibetan (
bo) - Training paradigm: Continual pretraining (CPT) starting from Gemma 3 PT 4B
Intended Use
This checkpoint is released primarily for research purposes. Researchers are welcome to use this CPT checkpoint as a base model for future work, particularly in model merging and logit fusion.
Citation
If you use this model, please cite:
@article{zhang2026efficient,
title={Efficient Low-Resource Language Adaptation via Multi-Source Dynamic Logit Fusion},
author={Zhang, Chen and Lin, Jiuheng and Liao, Zhiyuan and Feng, Yansong},
journal={arXiv preprint arXiv:2604.18106},
year={2026}
}