yurijmikhalevich/rclip-models

rclip model artifacts

This repository contains the inference artifacts downloaded by rclip. They are ONNX and Core ML format conversions of OpenCLIP's ViT-B-32-256 checkpoint datacomp_s34b_b86k; the learned model is not fine-tuned or otherwise changed.

Provenance and attribution

The source checkpoint is laion/CLIP-ViT-B-32-256x256-DataComp-s34B-b86K, trained by Mehdi Cherti on the DataComp-1B dataset using mlfoundations/open_clip. The tokenizer vocabulary originates from OpenAI CLIP through OpenCLIP.

See THIRD_PARTY_NOTICES.md for the complete attribution and LICENSE and OPENAI-CLIP-MIT.txt for the applicable MIT notices.

Files

• ViT-B-32-256-datacomp_s34b_b86k/visual.onnx: visual encoder for ONNX Runtime
• ViT-B-32-256-datacomp_s34b_b86k/textual.onnx: text encoder for ONNX Runtime
• ViT-B-32-256-datacomp_s34b_b86k/visual.mlpackage: visual encoder for Core ML
• tokenizer/bpe_simple_vocab_16e6.txt.gz: CLIP BPE vocabulary

Uses and limitations

The model supports image and text retrieval and zero-shot image classification. The upstream model card recommends task-specific testing and warns against untested deployment, surveillance, and facial-recognition uses. Review its full intended-use, limitation, bias, and training-data disclosures before using these artifacts outside rclip's local image-search workflow.

References

• DataComp paper
• OpenCLIP
• OpenAI CLIP