Vero-600K Multi-Task Visual Reasoning RL Dataset

## Overview Vero-600K is a curated reinforcement learning dataset containing approximately 600,000 samples drawn from 59 source datasets, spanning 6 diverse visual reasoning categories. It is distributed in Parquet format with image and text modalities, designed for training and evaluating multi-task visual reasoning in vision-language models (VLMs). The dataset is part of the Vero project, a fully open RL recipe from Princeton's Z-Lab. ## Schema - image — image — the visual input for the reasoning task - question/prompt — string — natural language query about the image - answer — string — ground-truth answer used as RL reward signal - category — string — one of 6 visual reasoning categories (STEM, Chart & OCR, Spatial & Action, etc.) - source_dataset — string — original dataset of provenance (among 59) - +additional metadata columns for RL training (See HF dataset page for the authoritative schema.) ## Sources - zlab-princeton/Vero-600k on Hugging Face — https://huggingface.co/datasets/zlab-princeton/Vero-600k — Apache-2.0 - Aggregates 59 upstream visual reasoning datasets (see HF dataset card for full provenance list) ## Methodology The Vero team curated 600K samples from 59 existing visual reasoning datasets, organizing them across 6 broad categories: STEM Reasoning, Chart & OCR, Spatial & Action, and others. Samples were selected and reformatted to support reinforcement learning with verifiable rewards for vision-language model training. Normalization includes standardizing prompt/answer formats across heterogeneous source datasets to enable unified multi-task RL training. ## Known gaps & limitations As an aggregation of 59 source datasets, quality and licensing of individual subsets may vary — buyers redistributing derivatives should verify upstream licenses for their use case. Coverage is English-only. The dataset is optimized for RL training signals rather than supervised fine-tuning, so reward verifiability may not equal answer correctness in all cases. Source does not document deduplication against common VLM evaluation benchmarks — buyers should validate empirically before using for benchmark-adjacent training. ## Intended use & out-of-scope - IS for: RL post-training of vision-language models, multi-task visual reasoning fine-tuning, evaluating VLM reasoning capabilities across STEM/Chart/OCR/Spatial domains. - IS NOT for: leakage-sensitive benchmark training (not deduplicated against MMMU/MathVista/etc.), non-English visual reasoning, or production use without per-subset license review. _Federated dataset: 978 parquet shards, 398.96 GB total. Queries and downloads stream through the DataBazaar API._ Original supplier listing: Vero-600K: Multi-Task Visual Reasoning RL Dataset 600K curated reinforcement learning samples from 59 datasets across 6 visual reasoning categories for training and evaluating vision-language models.