bn22/siglip_openclip_s_16_unsafe

A SigLIP-based ViT-S/16 vision-language model checkpoint hosted on Hugging Face under the bn22 namespace, designed for zero-shot image classification and text-image embedding via the OpenCLIP inference stack. It uses sigmoid loss pretraining (SigLIP objective) and is compatible with both OpenCLIP and timm workflows. Source: supabase.