ViRL39K: 38,870 Verifiable QAs for Vision-Language RL Training

ViRL39K contains 38,870 verifiable question-answer pairs designed for Vision-Language Reinforcement Learning training, released by TIGER-Lab in April 2025. It aggregates and refines data from seven specialized sources, including Llava-OneVision, MM-Math, and DeepScaleR, through a process of cleaning, reformatting, and verification.

Use Cases

• Training Vision-Language Reasoning Models using the verifiable QA pairs for reward signal generation
• Fine-tuning multimodal models on the MM-Math and MV-Math subsets for geometric and mathematical reasoning
• Implementing reinforcement learning pipelines using the reformatted chain-of-thought (M3CoT) data

Strengths

• 38,870 verifiable QA pairs
• Consolidates 7 distinct vision-language datasets including MM-Eureka and DeepScaleR
• MIT licensed for open research

Limitations

• Secondary dataset composed of reformatted existing sources rather than original primary collection
• Potential domain bias toward mathematical reasoning due to source selection (MM-Math, MV-Math)

Provenance

Source

TIGER-Lab (Arxiv 2504.08837)

Collection Method

Aggregated from existing datasets (Llava-OneVision, R1-OneVision, MM-Eureka, MM-Math, M3CoT, DeepScaleR, MV-Math) with cleaning and verification.

Freshness

Last updated April 2025.