NanoDB self-built dataset Embeddings

[JIA-HONG-CHU]

I want to use NanoDB to build my own data set embedding, but I found that
Use the same model ViT-L/14@336px
Use the same data set coco validation sets
However, the Embedding results created using the github main branch and the legacy branch are very different, and the results of the main branch are very poor. I would like to ask why there is such a gap between the two versions, or what is the reason why I cannot do this? Thank you