Vero-600k

Summary

Vero is a fully open reinforcement learning (RL) recipe for training and evaluating multi-task visual reasoning with vision-language models. This repository contains the **Vero-600K** dataset, a curation of 600K reinforcement learning samples from 59 datasets across 6 diverse visual reasoning categories.

Keywords

hf-dataset reinforcement-learning parquet image text datasets dask polars mlcroissant multimodal visual-reasoning has-paper

Topics

Reinforcement Learning

Research notes

• downloads=16497; likes=37