Fine-tune EmbeddingGemma

View on ai.google.dev Run in Google Colab Run in Kaggle Open in Vertex AI View source on GitHub

Fine-tuning helps close the gap between a model's general-purpose understanding and the specialized, high-performance accuracy that your application requires. Since no single model is perfect for every task, fine-tuning adapts it to your specific domain.

Imagine your company, "Shibuya Financial" offers various complex financial products like investment trusts, NISA accounts (a tax-advantaged savings account), and home loans. Your customer support team uses an internal knowledge base to quickly find answers to customer questions.

Setup

Before starting this tutorial, complete the following steps:

  • Get access to EmbeddingGemma by logging into Hugging Face and selecting Acknowledge license for a Gemma model.
  • Generate a Hugging Face Access Token and use it to login from Colab.

This notebook will run on either CPU or GPU.

Install Python packages

Install the libraries required for running the EmbeddingGemma model and generating embeddings. Sentence Transformers is a Python framework for text and image embeddings. For more information, see the Sentence Transformers documentation.

pip install -U sentence-transformers git+https://github.com/huggingface/transformers@v4.56.0-Embedding-Gemma-preview

After you have accepted the license, you need a valid Hugging Face Token to access the model.

# Login into Hugging Face Hub
from huggingface_hub import login
login()

Load Model

Use the sentence-transformers libraries to create an instance of a model class with EmbeddingGemma.

import torch
from sentence_transformers import SentenceTransformer

device = "cuda" if torch.cuda.is_available() else "cpu"

model_id = "google/embeddinggemma-300M"
model = SentenceTransformer(model_id).to(device=device)

print(f"Device: {model.device}")
print(model)
print("Total number of parameters in the model:", sum([p.numel() for _, p in model.named_parameters()]))

Device: cuda:0
SentenceTransformer(
  (0): Transformer({'max_seq_length': 2048, 'do_lower_case': False, 'architecture': 'Gemma3TextModel'})
  (1): Pooling({'word_embedding_dimension': 768, 'pooling_mode_cls_token': False, 'pooling_mode_mean_tokens': True, 'pooling_mode_max_tokens': False, 'pooling_mode_mean_sqrt_len_tokens': False, 'pooling_mode_weightedmean_tokens': False, 'pooling_mode_lasttoken': False, 'include_prompt': True})
  (2): Dense({'in_features': 768, 'out_features': 3072, 'bias': False, 'activation_function': 'torch.nn.modules.linear.Identity'})
  (3): Dense({'in_features': 3072, 'out_features': 768, 'bias': False, 'activation_function': 'torch.nn.modules.linear.Identity'})
  (4): Normalize()
)
Total number of parameters in the model: 307581696

Prepare the Fine-Tuning Dataset

This is the most crucial part. You need to create a dataset that teaches the model what "similar" means in your specific context. This data is often structured as triplets: (anchor, positive, negative)

  • Anchor: The original query or sentence.
  • Positive: A sentence that is semantically very similar or identical to the anchor.
  • Negative: A sentence that is on a related topic but semantically distinct.

In this example, we only prepared 3 triplets, but for a real application, you would need a much larger dataset to perform well.

from datasets import Dataset

dataset = [
    ["How do I open a NISA account?", "What is the procedure for starting a new tax-free investment account?", "I want to check the balance of my regular savings account."],
    ["Are there fees for making an early repayment on a home loan?", "If I pay back my house loan early, will there be any costs?", "What is the management fee for this investment trust?"],
    ["What is the coverage for medical insurance?", "Tell me about the benefits of the health insurance plan.", "What is the cancellation policy for my life insurance?"],
]

# Convert the list-based dataset into a list of dictionaries.
data_as_dicts = [ {"anchor": row[0], "positive": row[1], "negative": row[2]} for row in dataset ]

# Create a Hugging Face `Dataset` object from the list of dictionaries.
train_dataset = Dataset.from_list(data_as_dicts)
print(train_dataset)

Dataset({
    features: ['anchor', 'positive', 'negative'],
    num_rows: 3
})

Before Fine-Tuning

A search for "tax-free investment" might have given the following results, with similarity scores:

  1. Document: Opening a NISA account (Score: 0.45)
  2. Document: Opening a Regular Saving Account (Score: 0.48) <- Similar score, potentially confusing
  3. Document: Home Loan Application Guide (Score: 0.42)
task_name = "STS"

def get_scores(query, documents):
  # Calculate embeddings by calling model.encode()
  query_embeddings = model.encode(query, prompt=task_name)
  doc_embeddings = model.encode(documents, prompt=task_name)

  # Calculate the embedding similarities
  similarities = model.similarity(query_embeddings, doc_embeddings)

  for idx, doc in enumerate(documents):
    print("Document: ", doc, "-> 🤖 Score: ", similarities.numpy()[0][idx])

query = "I want to start a tax-free installment investment, what should I do?"
documents = ["Opening a NISA Account", "Opening a Regular Savings Account", "Home Loan Application Guide"]

get_scores(query, documents)

Document:  Opening a NISA Account -> 🤖 Score:  0.45698774
Document:  Opening a Regular Savings Account -> 🤖 Score:  0.48092696
Document:  Home Loan Application Guide -> 🤖 Score:  0.42127067

Training

Using a framework like sentence-transformers in Python, the base model gradually learns the subtle distinctions in your financial vocabulary.

from sentence_transformers import SentenceTransformerTrainer, SentenceTransformerTrainingArguments
from sentence_transformers.losses import MultipleNegativesRankingLoss
from transformers import TrainerCallback

loss = MultipleNegativesRankingLoss(model)

args = SentenceTransformerTrainingArguments(
    # Required parameter:
    output_dir="my-embedding-gemma",
    # Optional training parameters:
    prompts=model.prompts[task_name],    # use model's prompt to train
    num_train_epochs=5,
    per_device_train_batch_size=1,
    learning_rate=2e-5,
    warmup_ratio=0.1,
    # Optional tracking/debugging parameters:
    logging_steps=train_dataset.num_rows,
    report_to="none",
)

class MyCallback(TrainerCallback):
    "A callback that evaluates the model at the end of eopch"
    def __init__(self, evaluate):
        self.evaluate = evaluate # evaluate function

    def on_log(self, args, state, control, **kwargs):
        # Evaluate the model using text generation
        print(f"Step {state.global_step} finished. Running evaluation:")
        self.evaluate()

def evaluate():
  get_scores(query, documents)

trainer = SentenceTransformerTrainer(
    model=model,
    args=args,
    train_dataset=train_dataset,
    loss=loss,
    callbacks=[MyCallback(evaluate)]
)
trainer.train()

Step 3 finished. Running evaluation:
Document:  Opening a NISA Account -> 🤖 Score:  0.6449194
Document:  Opening a Regular Savings Account -> 🤖 Score:  0.44123
Document:  Home Loan Application Guide -> 🤖 Score:  0.46752414
Step 6 finished. Running evaluation:
Document:  Opening a NISA Account -> 🤖 Score:  0.68873787
Document:  Opening a Regular Savings Account -> 🤖 Score:  0.34069622
Document:  Home Loan Application Guide -> 🤖 Score:  0.50065553
Step 9 finished. Running evaluation:
Document:  Opening a NISA Account -> 🤖 Score:  0.7148906
Document:  Opening a Regular Savings Account -> 🤖 Score:  0.30480516
Document:  Home Loan Application Guide -> 🤖 Score:  0.52454984
Step 12 finished. Running evaluation:
Document:  Opening a NISA Account -> 🤖 Score:  0.72614634
Document:  Opening a Regular Savings Account -> 🤖 Score:  0.29255486
Document:  Home Loan Application Guide -> 🤖 Score:  0.5370023
Step 15 finished. Running evaluation:
Document:  Opening a NISA Account -> 🤖 Score:  0.7294032
Document:  Opening a Regular Savings Account -> 🤖 Score:  0.2893038
Document:  Home Loan Application Guide -> 🤖 Score:  0.54087913
Step 15 finished. Running evaluation:
Document:  Opening a NISA Account -> 🤖 Score:  0.7294032
Document:  Opening a Regular Savings Account -> 🤖 Score:  0.2893038
Document:  Home Loan Application Guide -> 🤖 Score:  0.54087913
TrainOutput(global_step=15, training_loss=0.009651281436261646, metrics={'train_runtime': 63.2486, 'train_samples_per_second': 0.237, 'train_steps_per_second': 0.237, 'total_flos': 0.0, 'train_loss': 0.009651281436261646, 'epoch': 5.0})

After Fine-Tuning

The same search now yields much clearer results:

  1. Document: Opening a NISA account (Score: 0.72) <- Much more confident
  2. Document: Opening a Regular Saving Account (Score: 0.28) <- Clearly less relevant
  3. Document: Home Loan Application Guide (Score: 0.54)
get_scores(query, documents)

Document:  Opening a NISA Account -> 🤖 Score:  0.7294032
Document:  Opening a Regular Savings Account -> 🤖 Score:  0.2893038
Document:  Home Loan Application Guide -> 🤖 Score:  0.54087913

To upload your model to the Hugging Face Hub, you can use the push_to_hub method from the Sentence Transformers library.

Uploading your model makes it easy to access for inference directly from the Hub, share with others, and version your work. Once uploaded, anyone can load your model with a single line of code, simply by referencing its unique model ID <username>/my-embedding-gemma

# Push to Hub
model.push_to_hub("my-embedding-gemma")

Summary and next steps

You have now learned how to adapt an EmbeddingGemma model for a specific domain by fine-tuning it with the Sentence Transformers library.

Explore what more you can do with EmbeddingGemma: