Skip to main content
Back

mrm8488's EmbeddingGemma 2 Fine-Tune Rivals 4B Models at 300M Parameters

A 0.3B EmbeddingGemma 2 fine-tune turns workflow states into typed, calibrated answers to yes/no, multiple-choice and ranked-level questions.

Model, Full Access
  • A 0.3B EmbeddingGemma 2 fine-tune that returns typed, calibrated answers to workflow decision questions.
  • Handles three primitives: choice, ordered score, and probabilistic yes/no noul.
  • Reaches 76.4% test accuracy on the typed-decisions benchmark, up from 67.7% with a frozen encoder.
  • Schema is plain data, so options and questions can be edited without retraining.
  • Apache 2.0, multilingual-capable, must run in bfloat16 or float32 (float16 NaNs silently).
  • Targets agent trace triage, customer service routing, invoice decisions and security alerts.

Manuel Romero, who publishes as mrm8488, has released embeddinggemma-2-ft-typed-decision, a 300-million-parameter fine-tune of Google’s EmbeddingGemma 2. Given a JSON-like workflow state and a question schema, the model returns temperature-calibrated probability distributions with typed values. The Apache 2.0-licensed checkpoint is available on Hugging Face and supports bfloat16 inference on a single GPU.

Repeated agent and business workflows often apply the same routing questions to every state. This model evaluates those questions in one call, preserves uncertainty for downstream thresholds, and removes free-text output parsing. Its parameter count is also substantially lower than that of multi-billion-parameter instruction models, although the release provides no latency or throughput comparison.

Typed questions from one state

Supported decision types
Type Purpose Example
choice Selects one labeled option from a schema. Continue, observe, review, or stop
score Assigns an ordered level from a rubric. Risk from 0 to 3
noul Returns a probabilistic Boolean decision. Escalate: true or false

Each example in the typed-decisions benchmark pairs one state with five questions. The benchmark describes this as a System One task, meaning bounded classification without multi-step generation. Every response retains the probability assigned to each allowed label.

The scoring head compares the text of every option with the state and question. Developers can rename options, revise descriptions, or add questions without retraining the checkpoint. Reliability can decline when a new schema differs substantially from the training distribution.

Semantic similarity, then correction

EmbeddingGemma 2 independently encodes the state, each question, and each option. Mean pooling averages the token representations into 768-dimensional vectors, which are then normalized to unit length. A schema-conditioned head of roughly 500,000 parameters combines those vectors.