Phi3.5-Vision - Data - training and post training

[estelleafl]

Hi!

Thank you for your work and for sharing it !
I wanted to know if the post-training data you used to do SFT (33B tokens) and the one you used for DPO were open source

Also, did you share the filtering scripts/methods you are referring to in the paper as well as the pretraining dataset?

Thank you

[leestott]

You can find detailed information about the Phi-3.5 models, including Phi-3.5 Vision, in the following academic papers and resources:

Phi-3 Technical Report on arXiv: This paper provides an in-depth look at the Phi-3.5 series, including the datasets and training methods used. It discusses the advancements in multilingual, multimodal, and long-context capabilities of the Phi-3.5 models.
https://arxiv.org/pdf/2404.14219

Microsoft Tech Community Blog: This blog post highlights the features and capabilities of the Phi-3.5 models, including Phi-3.5 Vision, and provides insights into their performance and applications. https://techcommunity.microsoft.com/blog/azure-ai-services-blog/discover-the-new-multi-lingual-high-quality-phi-3-5-slms/4225280

[estelleafl]

Thank you @leestott! Can you maybe point me to the data that was used if it was released?
I would like to reproduce Phi-3.5 Vision from the Phi3.5 language model. Is there a training script and data for that?

I could not find it

Best
Estelle