Layer-range L2-SP regularization: middle layers receive a larger L2 penalty against the base Gemma-3-4B weights than the first/last layers. Soft equivalent of layer freezing.

Base model: google/gemma-3-4b-pt
Strategy: layer-reg
Language family: Romance
Code: https://github.com/sanchit-ahuja/scaling-multilingual-experts

Loading

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("sanchitahuja205/xelm-gemma-4b-romance-layer-reg")
tokenizer = AutoTokenizer.from_pretrained("sanchitahuja205/xelm-gemma-4b-romance-layer-reg")

Training recipe

The exact training recipe lives in configs/yaml/train_gemma_layer_range.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_layer_range.yaml

Citation

bibtex

@misc{ahuja2026parameteralignmentmitigatescatastrophic,
      title={Parameter Alignment Mitigates Catastrophic Forgetting in Multilingual Expert Language Models},
      author={Sanchit Ahuja and Terra Blevins},
      year={2026},
      eprint={2606.00284},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2606.00284},
}