Starlings.Models.EmbeddingGemma2.MLNet 0.3.0

Starlings.Models.EmbeddingGemma2.MLNet

Text only. An ML.NET estimator and transformer that turn a text column into an EmbeddingGemma 2 embedding column (Vector<Single, 768>, or a Matryoshka size of 512, 256 or 128, L2-normalized) with the embedder in Starlings.Models.EmbeddingGemma2. The API may still change.

Maturity

Maturity is per profile, generated from the registry manifests by model-pipeline catalog (do not edit this table by hand). Gates, evidence and capabilities per profile: the model catalog.

Install

dotnet add package Starlings.Models.EmbeddingGemma2.MLNet
dotnet add package Microsoft.ML.OnnxRuntime --version 1.30.0          # or Microsoft.ML.OnnxRuntime.Gpu.Linux; exactly one

Use

using Starlings.Models.EmbeddingGemma2;
using Starlings.Models.EmbeddingGemma2.MLNet;
using Microsoft.ML;

var ml = new MLContext();
var files = EmbeddingGemma2ModelAssets.OpenLocal();      // offline, verified
// The pinned graph plus the owned tokenizer (Starlings.Tokenizers, profile gemma-bpe-embeddinggemma2).
using var transformer = ml.Transforms.ApplyEmbeddingGemma2Embedding(files, "Embedding", "Text",
        new EmbeddingGemma2EmbeddingOptions { Prompt = EmbeddingGemma2Prompt.Document, BatchSize = 8 })
    .Fit(data);
IDataView embedded = transformer.Transform(data);

• One inference per row, on the first output getter call; rows never read are never embedded.
• TokenCountColumnName (default none) adds an int column with each row's token count (special tokens and prompt included).
• Prompts are explicit (Prompt, default none). The dimension comes from Encoder.Dimension (768, 512, 256 or 128).
• BatchSize above 1 batches sequential, unshuffled cursors with a look-ahead cursor; row ids are checked.
• Saving the transformer throws (mlnet.persistence: unsupported); cursors take no cancellation token.

Status: supported on the Stable embeddinggemma-2/onnx-fp32-cpu profile (Linux x64 CPU; G7 passed; the contract is in the Starlings.Models.EmbeddingGemma2 README); supported over a CUDA session on the Stable embeddinggemma-2/onnx-fp32-cuda profile (Linux x64, RTX 3090; document rows within the CUDA profile's bound, rerun by the per-release GPU regression); Native AOT unsupported (Microsoft.ML).

License and usage policy

The package is MIT OR Apache-2.0. The model weights and Starling's ONNX graph derived from them are Apache-2.0 (Google DeepMind). The model card states:

Deployments must adhere to the Gemma Prohibited Use Policy.