Vero-2.5M-Unfiltered: Reinforcement Learning Samples for Vision-Language Models

2.5 million unfiltered reinforcement learning samples form the raw material for constructing the Vero-600k and Vero-1.6M datasets. The dataset, created by zlab-princeton, is intended for training multi-task visual reasoning models. It was last updated on June 11, 2026.

Use Cases

• Training multi-task visual reasoning models based on the described RL samples.
• Benchmarking model performance against unfiltered, raw training data.
• Analyzing the effects of question and answer filtering on dataset construction.
• Studying task category distribution and imbalance in large-scale RL datasets.

Strengths

• Contains 2.5 million reinforcement learning samples.
• Serves as the raw, unfiltered source for two other curated datasets (Vero-600k and Vero-1.6M).

Limitations

• Task categories are explicitly noted as not balanced.
• Description metadata is limited; actual data quality requires manual inspection after download.
• Column-level documentation is absent; field semantics must be inferred after download.

Provenance

Source

zlab-princeton

Collection Method

Curation of reinforcement learning samples.

Freshness

Last updated 2026-06-11 17:45:32.