This model is built on top of OLMo 2 1124 13B Instruct adapted for Nepali using 200M target language tokens sampled from MADLAD-400. The model is adapted using the HFT approach, a state-of-the-art static selective parameter update approach that updates exactly 50% of parameters using a fine-grained, per-layer strategy. Its freezing strategy is as follows: (1) for self-attention, it randomly freezes two of the four matrices ($W_Q, W_K, W_V, W_O$); (2) for feed-forward layers, it freezes two of three matrices ($W_{up}, W_{down}, W_{gate}$) in a random half of the layers and one matrix in the remaining half. This is based on https://aclanthology.org/2025.acl-long.626/.

Use the code below to get started with the model.

@misc{yamaguchi2025mitigatingcatastrophicforgettingtarget,
    title={Mitigating Catastrophic Forgetting in Target Language Adaptation of LLMs via Source-Shielded Updates}, 
    author={Atsuki Yamaguchi and Terufumi Morishita and Aline Villavicencio and Nikolaos Aletras},
    year={2025},
    eprint={2512.04844},
    archivePrefix={arXiv},
    primaryClass={cs.CL},
    url={https://arxiv.org/abs/2512.04844}, 
}