---
license: gemma
base_model: google/embeddinggemma-300m
pipeline_tag: feature-extraction
tags:
- coreml
- text-embeddings
- sentence-transformers
---

# EmbeddingGemma → Core ML

This repository contains Core ML conversions of Google's official
`embeddinggemma-300m` checkpoint. Each package uses a fixed 512-token input
and returns one normalized 768-dimensional embedding.

Conversion scripts, validation details, and the evidence trail live in the
[source repository](https://github.com/Mannyking/embeddinggemma-coreml).

## Packages

| Package | Minimum platform | Saved size |
| --- | --- | ---: |
| `f32-ios18/EmbeddingGemmaF32.mlpackage` | iOS/iPadOS 18, macOS 15 | 1,235.3 MB |
| `f16-ios15/EmbeddingGemmaF16.mlpackage` | iOS/iPadOS 15, macOS 12 | 620.4 MB |
| `int4-attention-int8-ios18/EmbeddingGemmaInt4AttentionInt8.mlpackage` | iOS/iPadOS 18, macOS 15 | 216.5 MB |

The F32 iOS 18 package is the source for the int4 candidate. The mixed-FP16
package uses FP16 only for `gather` and `linear` operations; attention,
normalization, pooling, and outputs remain Float32. The int4 candidate uses
per-block int4 linear weights, except attention Q/K/V/output projections,
which remain int8.

## Tokenizer and Inputs

Tokenizer assets are in [`tokenizer/`](./tree/main/tokenizer).

| Name | Type and shape |
| --- | --- |
| `input_ids` | Int32 `[1, 512]` |
| `attention_mask` | Int32 `[1, 512]` |
| `embedding` | Float32 `[1, 768]` |

Use right padding. Add BOS and EOS tokens; `<pad>` is ID 0, `<eos>` is ID 1,
`<bos>` is ID 2, and `<unk>` is ID 3. The 512-token limit includes the prompt
and special tokens.

Use these prompts before tokenization as adviced by The Gemma team:

- Query: `task: search result | query: `
- Document: `title: none | text: `

## Validation

The original checkpoint is pinned to
[`57c266a740f537b4dc058e1b0cda161fd15afa75`](https://huggingface.co/google/embeddinggemma-300m/tree/57c266a740f537b4dc058e1b0cda161fd15afa75).

The saved F32 iOS 18 package passed macOS assessment on ten fixtures,
including exact 511- and 512-token inputs; its largest recorded elementwise
error against the original-model baseline was `4.34e-7`.

The mixed-FP16 package passed its saved-artifact assessment with cosine
similarity from `0.999911` to `0.999960` against that baseline. The int4/int8
package is intentionally lossy; its exact 511- and 512-token fixtures had
cosine near `0.901` against F32. 

### BEIR SciFact Retrieval

The F32 and int4/int8 packages were evaluated across 252 eligible SciFact
claims and 4,799 abstracts. Records over the fixed 512-token limit were
excluded.

| Model | Saved size | Recall@1 | Recall@10 | nDCG@10 | MRR@10 |
| --- | ---: | ---: | ---: | ---: | ---: |
| Core ML F32 | 1,235.3 MB | 0.5742 | 0.8878 | 0.7488 | 0.7102 |
| Core ML int4 / attention int8 | 216.5 MB | 0.5671 | 0.8882 | 0.7414 | 0.7037 |

This is a quick, single-dataset retrieval check added to catch large quality
regressions from conversion or quantization. It is not a broad benchmark, an
app-quality claim, or a device-performance result.

## Source and License

Based on [google/embeddinggemma-300m](https://huggingface.co/google/embeddinggemma-300m).
These packages and tokenizer assets are Gemma model derivatives and are subject
to the [Gemma Terms of Use](https://ai.google.dev/gemma/terms).