Vero-1.6M: Multi-Task Visual Reasoning Samples for Reinforcement Learning

Vero-1.6M is an expanded reinforcement learning dataset containing 1.6 million samples curated from 59 source datasets for training vision-language models. It was created by zlab-princeton and last updated on June 11, 2026. The dataset is described as a fully open recipe for multi-task visual reasoning.

Use Cases

• Training multi-task visual reasoning agents based on the described reinforcement learning recipe.
• Evaluating vision-language model performance on a diverse set of visual tasks based on the 59 source datasets.
• Benchmarking reinforcement learning algorithms on a large-scale, curated collection of visual reasoning samples.

Strengths

• Contains 1.6 million reinforcement learning samples, expanded from a previous 600K version.
• Curated from a diverse set of 59 source datasets.
• Follows a documented and fully open reinforcement learning recipe.

Limitations

• Task categories are not balanced, which may introduce bias.
• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is known but other specifics like file formats and license are unknown.

Provenance

Source

zlab-princeton

Collection Method

Curated and filtered from 59 source datasets using a defined process.

Freshness

Last updated 2026-06-11 17:42:49; freshness should be verified.