LLaVA-OneVision 1.5 Multimodal Instruction Dataset

LLaVA-OneVision-1.5-Instruct is a 22 million instruction dataset curated by MVP-Lab for training large multimodal models. It was developed to support the LLaVA-OneVision-1.5 model family and was last updated in November 2025.

Use Cases

• Training instruction-following multimodal models using image-text instruction pairs.
• Fine-tuning large language models for visual question answering (VQA) tasks.
• Pretraining models on a mix of image captioning and other vision-language modalities.
• Benchmarking model performance on tasks derived from the 22M instruction examples.
• Studying data efficiency in multimodal learning with this curated collection.

Strengths

• Contains 22 million meticulously curated instruction examples.
• Designed to achieve state-of-the-art performance with reduced computational costs.

Limitations

• Specific column names, data structure, and sample composition are not publicly documented.
• The dataset's size and scope are described, but concrete sample data or file formats are unavailable for inspection.

Provenance

Source

MVP-Lab, associated with the LLaVA-OneVision-1.5 research project.

Collection Method

Collected and integrated during the development of the LLaVA-OneVision-1.5 model family.

Time Range

null

Freshness

Last updated on 2025-11 21.

Geography

null