LLaVA-OneVision-1.5: 85 Million Multimodal Mid-Training Records

Released by mvp-lab in 2025, this 85-million record multimodal collection supports the mid-training phase of the LLaVA-OneVision-1.5 framework. It aggregates image-text data from eight major sources including ImageNet-21k, LAIONCN, and SA-1B to facilitate democratized multimodal model training.

Use Cases

• Mid-stage training of multimodal models using the integrated image-text pairs
• Cross-domain visual feature extraction using the ImageNet-21k and COYO700M subsets
• Large-scale vision-language alignment using the MINT and Obelics text-image interleaved data

Strengths

• 85 million total records
• Aggregates 8 diverse source datasets including SA-1B and ImageNet-21k
• Apache 2.0 licensed for open research

Limitations

• Significant label noise inherent in web-scraped components like LAIONCN
• High computational overhead required for processing 85 million multimodal records

Provenance

Source

mvp-lab (Arxiv 2509.23661)

Collection Method

Aggregated from existing open-source datasets

Freshness

Last updated November 2024

Geography

Global