Pexels 568K InternVL2: Synthetic Image Captions for Multimodal Training

567,573 synthetic captions for images from the ptx0/photo-concept-bucket. The captions were generated using the OpenGVLab/InternVL2-40B-AWQ model and grounded using original image tags. The dataset was created by CaptionEmporium and last updated on Hugging Face in August 2024.

Use Cases

• Training text-to-image models based on the large-scale synthetic caption-image pairs.
• Fine-tuning vision-language models based on the captions generated by a large vision-language model.
• Benchmarking caption generation quality based on the synthetic outputs from InternVL2.
• Studying the properties of model-generated captions compared to human annotations.

Strengths

• Contains 567,573 synthetic captions, providing a substantial volume of training data.
• Captions were generated by a specific, advanced model (OpenGVLab/InternVL2-40B-AWQ).
• Captions were grounded using original image tags, which may improve relevance.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count for the source images is unknown, which may limit suitability assessment.
• The dataset consists of synthetic captions, which may contain model-specific biases or errors.

Provenance

Source

CaptionEmporium on Hugging Face.

Collection Method

Captions were synthetically produced using OpenGVLab/InternVL2-40B-AWQ for images from ptx0/photo-concept-bucket.

Freshness

Last updated 2024-08-21 22:39:24