Annoucement_blog_post

New Post: I recently adapted LeJEPA’s SIGReg for L₂-normalized text embeddings. Check out the blog post to see the hypersphere math behind the fix, and how it drastically reduces the batch size needed for contrastive learning. See here.