mihijito/embeddinggemma-300m-coreml-ios18

EmbeddingGemma-300M — CoreML (iOS 18+)

CoreML conversion of google/embeddinggemma-300m targeting iOS 18 (CoreML specification 9, opset CoreML8), so it loads on iOS 18.x devices where the earlier iOS 26-targeted conversion cannot.

Same export recipe as mlboydaisuke/embeddinggemma-300m-coreml (stateless bidirectional encoder, fixed 128-token sequence, int8 linear weight quantization, mean pool + dense head + L2 normalize, Matryoshka-truncatable output), rebuilt with coremltools 9.0 and minimum_deployment_target = iOS18. Embedding parity with the iOS 26 artifact: cosine ≥ 0.9999 at dim 256 across a mixed-length verification corpus, with no zero/NaN outputs for short (partial-mask) inputs.

Contents

• encoder.mlmodelc/ — compiled CoreML model (int8, seq len 128, fp16 I/O)
• model_config.json — I/O contract, task prefixes, Matryoshka dims
• hf_model/ — tokenizer files

I/O contract

• input_ids (1, 128) int32, zero-padded
• attention_mask (1, 128) fp16 — 1.0 valid, 0.0 pad
• → embedding (1, 768) fp16, L2-normalized; truncate + renormalize for 512/256/128

Prepend the task prefixes from model_config.json (e.g. "title: none | text: " for documents, "task: search result | query: " for queries) before tokenization.

License

Gemma is provided under and subject to the Gemma Terms of Use. By using this model you agree to those terms, including the Gemma Prohibited Use Policy. This repository redistributes a converted derivative of the original Google weights; all rights and restrictions of the original license apply.