Recaptioned Image-Text Dataset for Multimodal Alignment

Recap-DataComp-1B is a large-scale image-text dataset where descriptions have been enhanced using an advanced LLaVA-1.5-LLaMA3-8B model. The dataset was created by UCSC-VLAA and was last updated in January 2025.

Use Cases

• Training vision-language models on enhanced image-text pairs to improve caption generation.
• Benchmarking the performance of captioning models against recaptioned descriptions.
• Fine-tuning LLMs for tasks requiring detailed visual understanding using the recaptioned text.
• Studying the impact of advanced recaptioning on downstream multimodal tasks.

Strengths

• Dataset is recaptioned using an advanced LLaVA-1.5-LLaMA3-8B model.
• Based on the DataComp 1B scale, suggesting a foundation of 1 billion image-text pairs.

Limitations

• Specific row count, column details, and sample size are unknown.
• Potential bias inherited from the original DataComp 1B source data.
• The recaptioning model's specific error profile and its effect on label quality are not detailed.

Provenance

Source

UCSC-VLAA, based on the DataComp 1B dataset.

Collection Method

Recaptioning pipeline using a fine-tuned LLaMA-3-8B powered LLaVA-1.5 model.

Time Range

null

Freshness

Last updated on 2025-01 09.

Geography

null