Vero-600k

# Vero-600k <p align="center"> <img src="./vero-logo-blue-transparent.png" alt="Vero" width="520"> </p> 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. [](https://github.com/zlab-princeton/vero) [](https://huggingface.co/collections/zlab-princeton/vero) [](https://vero-reasoning.github.io) [](https://huggingface.co/papers/2604.04917) ## Highlights - **Scale**: 600K curated RL samples from 59 datasets. - **Diversity**: Covers 6 broad categories: STEM Reasoning, Chart & OCR, Spatial & Action, Knowledge & Recognition, Grounding & Counting, and Instruction Following. - **Task-Routed Rewards**: Designed to handle heterogeneous answer formats across diverse tasks. - **Open Recipe**: Fully open release of models, training code, evaluation suite, and dataset. ## Dataset Structure The dataset is organized into six broad task categories: 1. **STEM reasoning** 2. **Chart and OCR** 3. **Spatial reasoning and action** 4. **Knowledge and recognition** 5. **Grounding, counting, and visual search** 6. **Captioning and instruction following** For detailed dataset format, curation details, and reward routing metadata, see the [Data Guide](https://github.com/zlab-princeton/vero/blob/main/docs/DATA.md). ## Sample Usage To download and format the Vero-600k dataset for RL training using the official setup script: ```bash # Clone the repository git clone https://github.com/zlab-princeton/vero.git cd vero # Run the formatting script python scripts/download_and_format_vero_600k.py ``` This script exports images into `vero-rl/data/images/` and generates the training/validation `.verl.jsonl` files required for the Vero RL pipeline. ## Models Vero models are trained from various open-weight bases including Qwen3-VL, Qwen2.5-VL, and MiMo-VL. | Model | Base model | Params | | --- | --- | --- | | `Vero-Qwen3I-8B` | `Qwen3-VL-8B-Instruct` | 8B | | `Vero-Qwen3T-8B` | `Qwen3-VL-8B-Thinking` | 8B | | `Vero-MiMo-7B` | `MiMo-VL-7B-SFT` | 7B | | `Vero-Qwen25-7B` | `Qwen2.5-VL-7B-Instruct` | 7B | ## Citation If you use this dataset or the Vero recipe in your research, please cite: ```bibtex @article{sarch2026vero, title = {Vero: An Open RL Recipe for General Visual Reasoning}, author = {Sarch, Gabriel and Cai, Linrong and Wang, Qunzhong and Wu, Haoyang and Danqi Chen and Zhuang Liu}, year = {2026}, journal = {arXiv preprint arXiv:2604.04917}, } ``` ## License This dataset is released under the [Apache License 2.0](LICENSE). Users should also review the licenses and usage terms of the underlying base models and any upstream datasets included in this curation.