zlab-princeton

zlab-princeton / vero

Public

Vero: An Open RL Recipe for General Visual Reasoning

vero-reasoning.github.io
49
3
100% credibility
Found Apr 08, 2026 at 49 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
Python
AI Summary

Vero is an open-source toolkit for training and evaluating AI models on diverse visual reasoning tasks like charts, math diagrams, spatial understanding, and object recognition using reinforcement learning.

How It Works

1
🔍 Discover Vero

You stumble upon Vero, a friendly toolkit from university researchers that helps create smart AI assistants for understanding pictures, solving math from images, reading charts, and spotting objects.

2
🛠️ Prepare your workspace

Follow easy steps to set up your computer so it's ready for building visual AI helpers.

3
📥 Gather teaching examples

Download a big collection of 600,000 real-world image questions and answers to train your AI.

4
🚀 Start AI training

Hit go on a training session where your AI learns to reason about visuals like charts, shapes, and instructions using smart rewards.

5
🧪 Test your creation

Run quick checks on 30 fun challenges to see how well your AI handles visual puzzles across science, maps, and more.

🎉 Enjoy smart results

Your AI now shines at visual tasks – celebrate with impressive scores and ready-to-use helpers for everyday picture smarts!

Sign up to see the full architecture

4 more

Sign Up Free

Star Growth

See how this repo grew from 49 to 49 stars Sign Up Free
Repurpose This Repo

Repurpose is a Pro feature

Generate ready-to-use prompts for X threads, LinkedIn posts, blog posts, YouTube scripts, and more -- with full repo context baked in.

Unlock Repurpose
AI-Generated Review

What is vero?

Vero provides a complete Python-based recipe for training and evaluating vision-language models on general visual reasoning tasks, tackling the lack of open, scalable RL pipelines for multimodal AI. It delivers 600K curated samples across STEM, charts, spatial reasoning, and more, plus a 30-benchmark eval suite covering everything from OCR to instruction following. Users get pretrained checkpoints on models like Qwen2.5-VL, ready-to-run training scripts, and a simple eval harness—think a vero recipe book for Python visual reasoning devs.

Why is it gaining traction?

Unlike fragmented VLM toolkits, Vero bundles single-stage RL with task-routed rewards and LLM judging in one open stack, supporting Qwen, MiMo, and others out of the box. Devs love the bash-driven setup, data download script, and domain-specific eval commands like `bash eval_domain.sh --domain chart_ocr`. Its broad 59-dataset coverage and veRL/lmms-eval foundations make it a plug-and-play alternative to proprietary fine-tuning.

Who should use this?

Multimodal researchers benchmarking VLMs on real-world tasks like ChartQA or MMMU-Pro, or engineers at startups fine-tuning Qwen VL for apps needing chart parsing, spatial actions, or visual search. Ideal for teams handling general visual reasoning without building RL infra from scratch, especially those eyeing open models like Vero-Qwen3I-8B.

Verdict

Grab Vero if you're in visual reasoning—strong docs, pretrained models, and eval suite make it usable now despite 49 stars and 1.0% credibility signaling early maturity. Test with the quickstart training script; scale once production-ready.

(198 words)

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.