MegaPairs: 26 Million Multimodal Retrieval Instruction-Tuning Triplets

MegaPairs is a dataset containing over 26 million multimodal retrieval instruction-tuning triplets for AI/ML practitioners. The dataset was created by JUNJIE99 and uploaded to Hugging Face on 2025-04-13. Images in the dataset have been resized to a resolution of 512 × 512 to enhance accessibility.

Use Cases

• Fine-tuning models for cross-modal retrieval based on instruction-following triplets.
• Training models to align images and text based on large-scale synthetic data.
• Benchmarking universal multimodal retrieval systems on a massive scale.

Strengths

• Contains over 26 million instruction-tuning triplets, indicating a large scale.
• Images are uniformly resized to 512 × 512 resolution for consistency and accessibility.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.

Provenance

Source

JUNJIE99 on Hugging Face

Collection Method

Synthesized for multimodal retrieval instruction-tuning.

Freshness

Last updated 2025-04-12 17:41:52; freshness should be verified.