Published November 22, 2024 | Version v2
Figure Open

Scaling plots for 80% and 100% text-pretrained

Authors/Creators

Description

Scaling plots for 80% and 100% (text80 and text100) in the name
All are text-pretrained with 4 epochs of fine-tuning and 28b of image-text training with 10% image caption and 90% text (ratio 0p1 in the name)
plots with various ablations at 154m, 411m, 1b scale (appended in the name)

The text data is DCLM dataset and the image data is DatacompDR

Files

finetuning_epochs_datacompdronly_vs_vision_textstable-score_grid_ratio0p1_text80_154m.png

Additional details

Identifiers

Other
Figures_rebuttal_ICLR_2025

Dates

Submitted
2024-11-22