mlfoundations / clip_quality_not_quantity Goto Github PK

Great work, love this direction! Any chance you'll make the trained weights available? I am interested in doing further comparative analysis for research purposes

Hey, thanks for the great work and releasing code for accessing the exact datasets used for your analysis. I noticed that there was an earlier closed issue that asked about releasing the pre-trained models used, are there any further updates on the models release?
Further, since your work uses CLIP-RN50 models for all your releases, I was wondering if you are aware of any other work that does an analysis with vision-transformers for different data distributions at the medium data scale (<20M), specifically on the CC12M dataset/YFCC15M dataset? I am aware of the Datacomp medium-scale models (12.8M samples) but however I am mainly interested in accessing pre-trained models on the CC12M/YFCC15M dataset.
Looking forward to hearing your thoughts, thanks!