@version 1
# Label models: image-text encoders that place images and text in one space, so an image can be
# scored against any phrase. Each [model NAME] names the files it needs from this folder
# (fetch_models.jai downloads them) and how it is run; labels_create takes the NAME.
#
#   title                    shown to people
#   vision, text, tokenizer  file names, quoted: the image encoder, the text encoder, tokenizer.json
#   image_size               images are resized to this square for the vision model (px)
#   max_tokens               text is cut or padded to this many tokens
#   logit_scale, logit_bias  a phrase's score for an image is sigmoid(e^logit_scale * cosine + logit_bias),
#                            from the trained model (they are not in the ONNX encoders)
#   image_batch              images run through the vision model together
#   vision_coreml            on macOS, the vision model as a Core ML package, quoted, run on the GPU
#                            (tools/export_siglip_coreml.py makes it); without it, onnxruntime runs it

[model siglip2-base-patch16-384]
title:       "SigLIP 2 base, 384 px"
vision:      "siglip2-base-patch16-384-vision.onnx"
text:        "siglip2-base-patch16-384-text-int8.onnx"
tokenizer:   "siglip2-tokenizer.json"
image_size:  384
max_tokens:  64
logit_scale: 4.726024
logit_bias:  -16.772243
image_batch: 8
vision_coreml: "siglip2-base-patch16-384-vision.mlpackage"