EmbeddingGemma-300M Fine-tuned for Biblical Text Search (ONNX)

This is the ONNX version of our fine-tuned EmbeddingGemma-300M model specialized for biblical text search and retrieval. This version is optimized for web deployment using transformers.js.

Model Performance

Accuracy@1: 12.00% (13x improvement over base model)
Accuracy@3: 15.00%
Accuracy@10: 31.00%
Training Steps: 25 (optimal stopping point)
Base Model Accuracy@1: 0.91%

Usage with Transformers.js

import { AutoTokenizer, AutoModel } from '@huggingface/transformers';

// Load the model
const model = await AutoModel.from_pretrained('dpshade22/embeddinggemma-scripture-v1-onnx');
const tokenizer = await AutoTokenizer.from_pretrained('dpshade22/embeddinggemma-scripture-v1-onnx');

// Encode queries (use search_query: prefix)
const query = "search_query: What is love?";
const query_embedding = await model.encode([query]);

// Encode documents (use search_document: prefix)  
const document = "search_document: Love is patient and kind";
const doc_embedding = await model.encode([document]);

Prefixes

For optimal performance, use these prefixes:

Queries: "search_query: your question here"
Documents: "search_document: scripture text here"

Model Details

Base Model: google/embeddinggemma-300m
Training Data: 26,276 biblical text pairs
Training Steps: 25 steps (optimal stopping point)
Learning Rate: 2.0e-04
Batch Size: 8
Output Dimensions: 768D (supports Matryoshka 384D, 128D)
ONNX Conversion: Using nixiesearch/onnx-convert specialized tool

Training Details

Training Data: 26,276 biblical text pairs
Learning Rate: 2.0e-04
Batch Size: 8
Training Strategy: Early stopping at 25 steps to prevent overfitting
Output Dimensions: 768D (supports Matryoshka 384D, 128D)

Intended Use

This model is designed for:

Biblical text search and retrieval in web applications
Finding relevant scripture passages
Semantic similarity of religious texts
Question answering on biblical topics
Offline PWA applications using transformers.js

Conversion Details

Converted using: nixiesearch/onnx-convert specialized tool
ONNX Opset: 17
Optimization Level: 1
Max difference from original: 1.9e-05 (within acceptable tolerance)

Related Models

Original PyTorch version: dpshade22/embeddinggemma-scripture-v1
Base model: google/embeddinggemma-300m
Reference ONNX: onnx-community/embeddinggemma-300m-ONNX