Code used to generate the dataset

[andimarafioti]

Hi! Thank you for the contribution with the dataset! Really cool stuff! I was wondering, are you planing to release the code you used to create the dataset?

[ImKeTT]

Thank you for your interest in our work! We'll release the pipeline for recaptioning the dataset soon.
Also, we have already released our recaption model (the LLaMA3-powered LLaVA) here: https://huggingface.co/tennant/llava-llama-3-8b-hqedit

[pbaylies]

Thank you for releasing the recaptioning model weights; did you run it using Transformers, or with the original Llava repo? I tried it in transformers, and the transformers library complains about a missing preprocessor_config.json file, and also it notes that the model type is set as llava_llama which it does not recognize.

[ImKeTT]

Thanks for your interest @pbaylies !
We used a slightly modified version of the original LLaVA repo on GPU (we changed the conversation template to LLaMA3's) and a Jax-implemented version on TPU for inference. We'll release both inference pipelines in a few days, so stay tuned!