xelm-gemma-4b-austronesian-window-10-16-l2sp

Sliding-window continual pre-training with layer-range L2-SP regularization against the base model weights.

• Base model: google/gemma-3-4b-pt
• Strategy: layer_range_l2sp
• Language family: Austronesian
• Code: scaling-multilingual-experts

Loading

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("sanchitahuja205/xelm-gemma-4b-austronesian-window-10-16-l2sp")
tokenizer = AutoTokenizer.from_pretrained("sanchitahuja205/xelm-gemma-4b-austronesian-window-10-16-l2sp")

Training recipe

The exact training recipe lives in configs/yaml/train_gemma_early_window_l2sp.yaml in the code repo. The resolved config used for this specific run is also included in this model repo as training_config.yaml — load it with pyrallis to reproduce the run bit-for-bit:

python train.py --config_path configs/yaml/train_gemma_early_window_l2sp.yaml

Citation

@misc{ahuja2026specializingforgettinganalyzingknowledge,
      title={Specializing Without Forgetting: Analyzing Knowledge Preservation in Multilingual Model Adaptation},
      author={Sanchit Ahuja and Terra Blevins},
      year={2026},
      eprint={2606.00284},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2606.00284},
}