Modelpublic

sanchitahuja205/xelm-gemma-4b-dense

sourceHugging Facegemmaupdated 4mo agoView on Hugging Face

0likes7downloads

Model Card

xelm-gemma-4b-dense

Dense continual pre-training (CPT) of Gemma-3-4B on a concatenated 25B-token mixture across Slavic, Germanic, Indic, Austronesian, and Romance language families. This is the no-regularization baseline.

—Base model: google/gemma-3-4b-pt
—Strategy: dense
—Language families: all 5 (Slavic, Germanic, Indic, Austronesian, 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-dense")
tokenizer = AutoTokenizer.from_pretrained("sanchitahuja205/xelm-gemma-4b-dense")

Training recipe

The exact training recipe lives in `configs/yaml/train_gemma_dense.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:

bash

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

Reproducing the dense-reverted variant

bash

python train.py --config_path configs/yaml/revert_gemma_checkpoint.yaml \
    --revert.checkpoint_path $(huggingface-cli download sanchitahuja205/xelm-gemma-4b-dense) \
    --revert.revert_output_path ./reverted

Reproducing the expert soup (uniform average of 5 experts)

bash

python model_soup.py \
    --experts sanchitahuja205/xelm-gemma-4b-slavic-expert \
              sanchitahuja205/xelm-gemma-4b-germanic-expert \
              sanchitahuja205/xelm-gemma-4b-indic-expert \
              sanchitahuja205/xelm-gemma-4b-austronesian-expert \
              sanchitahuja205/xelm-gemma-4b-romance-expert \
    --output_dir ./xelm-gemma-4b-expert-soup \
    --alpha 1.0

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},
}