Conceptual Captions Cc12m Llavanext: 21.9M Synthetic Image Captions

A collection of 21,930,344 synthetic English captions for 10,965,172 images from the conceptual_12m dataset. The captions were generated using the llama3-llava-next-8b model, followed by cleanup and shortening with Meta-Llama-3-8B. The dataset was created by CaptionEmporium and last updated on Hugging Face in June 2024.

Use Cases

• Training vision-language models based on the large-scale synthetic caption-image pairs.
• Benchmarking image captioning models against LLM-generated descriptions.
• Fine-tuning text-to-image generation models using the provided image-text associations.
• Studying the properties and biases of synthetic captions produced by the LLaVA-Next architecture.

Strengths

• Contains 21.9 million synthetic captions, providing a large-scale resource.
• Pairs captions with 10.96 million images from the established conceptual_12m source.
• Captions were generated and refined using specific, named LLMs (llama3-llava-next-8b, Meta-Llama-3-8B).

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Data may reflect biases inherent to the source images and the caption-generating models.
• Row count is unknown, which may limit suitability assessment.

Provenance

Source

CaptionEmporium on Hugging Face, derived from the conceptual_12m dataset.

Collection Method

Captions were synthetically generated using llama3-llava-next-8b and Meta-Llama-3-8B models.

Freshness

Last updated 2024-06-30 03:04:24; freshness should be verified.