KALAVAI — Fiction Specialist (Qwen2.5-1.5B, seed 42)

Fine-tuned Qwen/Qwen2.5-1.5B on Fiction data as part of the KALAVAI decentralized cooperative training protocol.

Code: https://github.com/mechramc/Kalavai

Paper results

Qwen-2.5-1.5B specialists. MoE fusion: +1.06% ±0.01pp over best specialist (3 seeds). Mean divergence 3.16% — near floor of gain-divergence relationship.

How to use

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("mechramc/kalavai-qwen-fiction-specialist-seed42")
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-1.5B")

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}
}