Open Replication Matches Specialized Jev Decision Model with Simple LoRA Fine-tuning

A month-long open replication of the Jev-style decision model shows that just 400 steps of LoRA fine-tuning on an open model, with options labeled A/B/C, scores 84.9 and outperforms the full specialized system.

2026-10-24 ~ 2026-10-24 · 2 related posts

• Ablating Jev-style decision models: LoRA-only hits 84.9, beating the full recipe — vanstriendaniel · 2026-10-24
• 400 LoRA steps on Gemma 4 31B beat dedicated decision model Jev — vanstriendaniel · 2026-10-24