fuvty's picture
Update README.md
dad0db8 verified
metadata
library_name: r2r
pipeline_tag: text-classification
tags:
  - router
  - efficiency
  - language-model
datasets:
  - nics-efc/R2R_Router_Training
  - nics-efc/R2R_query
language:
  - en

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.

They are the routers from paper R2R: Efficiently Navigating Divergent Reasoning Paths with Small-Large Model Token Routing

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: