pairmuse/marqo-fashionSigLIP
A mirror of Marqo/marqo-fashionSigLIP
at upstream revision c56244cc94f92419e8369fa71efdaf403b124ce8, published by
PairMuse so that the checkpoint its image-embedding sidecar executes is pinned
to bytes it has reviewed. The weights, the open_clip configuration, the
preprocessor and the tokenizer files are the upstream files byte for byte.
Two files differ from upstream, both reviewed in the PairMuse repository
(defaults/models/marqo-fashionsiglip/):
marqo_fashionSigLIP.py:MarqoFashionSigLIPConfigdeclaresmodel_type = "siglip", the valueconfig.jsonalready carries, so transformers no longer logs a model-type mismatch on every load.config.json:open_clip_model_namepoints at this mirror, so open_clip reads the weights from the same pinned repository.
The model is Marqo's work: a SigLIP ViT-B-16 fine-tuned for fashion retrieval (up to 57 % better MRR and recall than FashionCLIP on their benchmarks). Please cite and credit Marqo; see the upstream model card for the evaluation and the training details. Licence: Apache-2.0, as upstream.