R2R_router_collections

R2R Router Models

This repository provides a collection of R2R routers (Mixture of Small and Large Language Models) and its training config built for different model pairs.

Model Description

R2R routers are lightweight classifiers that decide, at the token level, whether to generate with a small language model (SLM) or delegate to a large language model (LLM). The goal is to retain LLM-level quality while improving end-to-end efficiency.

We currently support routers for the Qwen3 series and the DeepSeek-R1-Qwen series under deterministic (non-sampling) decoding. In addition, we provide a router tailored for routing between DeepSeek-R1-Qwen-1.5B and DeepSeek-R1-Qwen-32B under DeepSeek’s default sampling settings(temperature=0.6, top_p=0.95).

Usage

For setup instructions, checkpoints, and examples, please visit our GitHub repository:

• GitHub: https://github.com/thu-nics/R2R
• Project page: https://fuvty.github.io/R2R_Project_Page/