EmbeddingGemma 300M, compact format for Omoikane

This is google/embeddinggemma-300m converted to the format Omoikane loads (omoi model install --semantic downloads it from here). It is not a new model and it was not trained or tuned: the weights are Google's, stored differently.

• tokens.i8 and tokens.escala: the 262,144 × 768 token embedding table in int8, one float32 scale per row (symmetric, max/127). Omoikane maps it from disk and reads only the rows of the tokens it sees.
• pesos.safetensors: every other weight in float16, including the two dense layers of the sentence-transformers head (densa1.weight, densa2.weight).
• config.json and tokenizer.json: unchanged.

Vectors computed from this format match the original model through sentence-transformers with cosine 0.999 or better, up to 2,048 tokens. The conversion is omoikane::gemma::compactar in Omoikane's source.

License

Gemma is provided under and subject to the Gemma Terms of Use found at ai.google.dev/gemma/terms. Its use is restricted by the Gemma Prohibited Use Policy, and by downloading or using these files you agree to both. This is a Model Derivative under those terms, distributed with the same restrictions; see NOTICE.