CoreML model seems wrong (dimension is 128 not 512)
#9
by
jens-alfke
- opened
I have doubts about the CoreML model in the repo.
The description of gte-small says the max token length is 512, but the metadata of the CoreML model says the input has dimension 128, which would make it impossible to pass more than 128 tokens.
Has anyone used this model successfully?
Hi, can you give the link to the model above?
It's the one checked into the repo, coreml/feature-extraction/float32_model.mlpackage.
It appears to be the correct model -- once I got it working in an app, it produces identical outputs to GTE-Small running in PyTorch. But I can't give it more than 128 input tokens or CoreML returns an error.