KALAVAI — Welsh Specialist (pythia-410m, seed 137)

Fine-tuned EleutherAI/pythia-410m on Welsh data as part of the KALAVAI decentralized cooperative training protocol.

Paper: KALAVAI: Predicting When Independent Specialist Fusion Works -- A Quantitative Model for Post-Hoc Cooperative LLM Training
GitHub: https://github.com/mechramc/Kalavai

Paper results

Yoruba PPL 41.9→7.7 (5.4×), Welsh 102.7→22.1 (4.6×), Tamil 4.2→3.0. MoE fusion of 4 specialists: +21.76% over best specialist (seeds 137+2026).

How to use

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("mechramc/kalavai-cross-lingual-welsh-specialist-seed137")
tokenizer = AutoTokenizer.from_pretrained("EleutherAI/pythia-410m")

This model is one specialist in a KALAVAI cooperative. To reproduce the MoE fusion results from the paper, load multiple domain specialists and combine them with a trained MoE router (see the paper and GitHub for details).

Citation

@article{kumaresan2026kalavai,
  title     = {{KALAVAI}: Predicting When Independent Specialist Fusion Works
               --- A Quantitative Model for Post-Hoc Cooperative {LLM} Training},
  author    = {Kumaresan, Ramchand},
  journal   = {arXiv preprint arXiv:2603.22755},
  year      = {2026},
  url       = {https://arxiv.org/abs/2603.22755}
}