How to Use generate_text_prompts for Custom Dataset Embeddings and Reparameterization Fine-tuning

[TAKAGNU]

I am trying to use the generate_text_prompts function to generate text embeddings for a custom dataset in a format similar to Flickr30k. My goal is to use these embeddings to reparameterize and fine-tune YOLO-World. Here is my current workflow:

I am using the tokens_positive_eval corresponding text as the input categories for generate_text_prompts to generate the text embeddings.

However, this results in a large number of categories (73,350), which requires setting num_classes and num_training_classes to 73,350 during reparameterization fine-tuning.

As a beginner, I am unsure if this is the correct approach. Specifically, I have the following questions:

Is it appropriate to use tokens_positive_eval text as the input for generate_text_prompts to generate embeddings for a custom dataset?

Is setting num_classes and num_training_classes to 73,350 the correct way to handle such a large number of categories during fine-tuning?

Are there any best practices or recommended workflows for generating text embeddings and fine-tuning YOLO-World on custom datasets with a large number of categories?

Any guidance or suggestions would be greatly appreciated!

Thank you in advance for your help!