cgerdes/embeddinggemma-2-coreml-mirror

The EmbeddingGemma 2 search model the Apiscript app downloads, pinned so a deleted or rewritten upstream cannot break installs.

Generated by scripts/package-embeddinggemma.sh — do not hand-edit.

Upstream: google/embeddinggemma-2 (Google, Apache-2.0) at revision 914f7f89142e33e77833254d9c9b90c3cef7303b. The text encoder only was converted to Core ML: 8-bit per-channel weights, fp16 compute, fixed input of 512 tokens (input_ids and attention_mask, int32 [1, 512]), output embedding fp32 [1, 768], mean-pooled and L2-normalised. Use the official prompts: task: search result | query: for queries, title: <title or none> | text: for documents. Every vector matched sentence-transformers fp32 at cosine >= 0.99 on 50 German and English inputs before upload. LICENSE and NOTICE are also inside the archive; the tokenizer files are byte-identical to upstream.

file	sha256	bytes
embeddinggemma-2-text-512-w8.aar	abcc012b7771732d043993c6c540c160d18f3cd457a3dd3e86b0b3d36364ebae	249615861
tokenizer.json	4d777ef5bdc1aa36227abdfb77c3e49e7b9c892d16e1b6bda41c393504828be4	32170510
tokenizer_config.json	17bd5d6e9364ca49a534e1502076593317c298d4a663623091ed45388f004874	1599