Bitfarmy/trama-embeddinggemma

EmbeddingGemma 300M for LiteRT (copy used by the Trama browser)

This repository hosts an unmodified copy of two files published by Google in `litert-community/embeddinggemma-300m`, so that the Trama Android browser can download them, only when the user explicitly enables on-device semantic search. Everything then runs on the phone; no text ever leaves it.

The files have not been modified in any way: same bytes as the originals.

License and use restrictions

Gemma is provided under and subject to the Gemma Terms of Use found at ai.google.dev/gemma/terms.

• —A full copy of the terms is included in this repository: `GEMMA_TERMS_OF_USE.md`.
• —The use restrictions of section 3.2 and the Gemma Prohibited Use Policy apply to anyone who downloads or uses these files. A copy of the policy is included: `GEMMA_PROHIBITED_USE_POLICY.md`.
• —By downloading or using these files you agree to the Gemma Terms of Use and to the Gemma Prohibited Use Policy, which are binding on you.
• —See also the `NOTICE` file.

Technical notes

• —Model: EmbeddingGemma 300M (Google DeepMind), mixed-precision quantization (int4 embeddings/feed-forward/projection, int8 attention), for the LiteRT runtime.
• —Signature embed_256: input text_batch int32 [1, 256] (token ids), output encodings float32 [1, 768].
• —Input preparation (as in Google's official LiteRT sample): SentencePiece tokens, truncated to 254, BOS first, EOS last, padded with PAD to 256.
• —Task prefixes from the official model card: task: search result | query: for queries, title: none | text: for documents.
• —Output: 768 dimensions; Trama keeps the first 256 (Matryoshka) and re-normalizes them.