single-cell/rbio

Overview

rbio1 is a 3B biology reasoning model from the Chan Zuckerberg Initiative, post-trained from Qwen2.5-3B-Instruct by reinforcement learning (GRPO) with virtual-cell models and the Gene Ontology as soft verifiers. It comes in about ten variants, one per verifier. The weights are distributed by CZI's Virtual Cells Platform, not on Hugging Face; this repository holds the code.

Test this model

Not servable. No inference endpoint is declared for this model, so there is nothing to call — no button here and no API route for it. What makes a model servable

Papers

• rbio1 - training scientific reasoning LLMs with biological world models as soft verifiersReference
  bioRxiv (Cold Spring Harbor Laboratory), 2025DOI: 10.1101/2025.08.18.670981 ↗