Skip to content

R1-V

Train visual counting generalisation at low cost

Visit official site

What R1-V is

The project is described as an effort to improve visual language model generalisation through reinforcement learning with verifiable rewards.

What you can do with it

  • Apply RLVR to visual counting tasks
  • Train a 2B model within 100 steps for OOD tests
  • Use open source code, models and datasets

Getting started

  • Clone the repository to a local machine
  • Install the required Python packages
  • Set environment variables such as DEBUG_MODE and LOG_PATH
  • Launch training with torchrun and the output, model and dataset paths

Pros & cons

✓ What we like

  • RLVR is stated to outperform CoT-SFT on generalisation
  • Multiple training configurations fit different hardware

! What to watch out for

  • Training needs a torchrun command with set paths
  • Environment variables such as DEBUG_MODE are required

FAQ

What is the stated training cost?

The material states 2.62 USD for 30 minutes on 8 A100 GPUs.

What is released?

The feature list names complete open source code, models and datasets.

Last reviewed: 2026-11-15

More AI coding tools tools

View all →

How we review