gemma4-research/tooling/fine-tuning/huggingface-recipes/README.md

Hugging Face Gemma Recipes

🤗💎 Welcome! This repository contains minimal recipes to get started quickly with the Gemma family of models.

Note

Gemma 4 Multimodal inference (vision, video, audio, function calling, object detection):

Getting Started

To quickly run a Gemma 💎 model on your machine, install the latest version of timm (for the vision encoder) and 🤗 transformers to run inference, or if you want to fine tune it.

$ pip install -U -q transformers timm

Inference with pipeline

The easiest way to start using Gemma 3n is by using the pipeline abstraction in transformers:

import torch
from transformers import pipeline

pipe = pipeline(
   "image-text-to-text",
   model="google/gemma-3n-E4B-it", # "google/gemma-3n-E4B-it"
   device="cuda",
   torch_dtype=torch.bfloat16
)

messages = [
   {
       "role": "user",
       "content": [
           {"type": "image", "url": "https://huggingface.co/datasets/ariG23498/demo-data/resolve/main/airplane.jpg"},
           {"type": "text", "text": "Describe this image"}
       ]
   }
]

output = pipe(text=messages, max_new_tokens=32)
print(output[0]["generated_text"][-1]["content"])

Detailed inference with transformers

Initialize the model and the processor from the Hub, and write the model_generation function that takes care of processing the prompts and running the inference on the model.

from transformers import AutoProcessor, AutoModelForImageTextToText
import torch

model_id = "google/gemma-3n-e4b-it" # google/gemma-3n-e2b-it
processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForImageTextToText.from_pretrained(model_id).to(device)

def model_generation(model, messages):
    inputs = processor.apply_chat_template(
        messages,
        add_generation_prompt=True,
        tokenize=True,
        return_dict=True,
        return_tensors="pt",
    )
    input_len = inputs["input_ids"].shape[-1]

    inputs = inputs.to(model.device, dtype=model.dtype)

    with torch.inference_mode():
        generation = model.generate(**inputs, max_new_tokens=32, disable_compile=False)
        generation = generation[:, input_len:]

    decoded = processor.batch_decode(generation, skip_special_tokens=True)
    print(decoded[0])

And then using calling it with our specific modality:

Text only

# Text Only

messages = [
    {
        "role": "user",
        "content": [
            {"type": "text", "text": "What is the capital of France?"}
        ]
    }
]
model_generation(model, messages)

Interleaved with Audio

# Interleaved with Audio

messages = [
    {
        "role": "user",
        "content": [
            {"type": "text", "text": "Transcribe the following speech segment in English:"},
            {"type": "audio", "audio": "https://huggingface.co/datasets/ariG23498/demo-data/resolve/main/speech.wav"},
        ]
    }
]
model_generation(model, messages)

Interleaved with Image/Video

# Interleaved with Image

messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "image": "https://huggingface.co/datasets/ariG23498/demo-data/resolve/main/airplane.jpg"},
            {"type": "text", "text": "Describe this image."}
        ]
    }
]
model_generation(model, messages)

Inference

Gemma 4

Notebooks

• Multimodal inference with Gemma 4 (vision, video, audio, function calling, object detection)

Gemma 3n

Notebooks

• Multimodal inference using Gemma 3n via pipeline

Function Calling

Gemma 3n

Notebooks

• Function Calling with Gemma 3n: Local File Reader

Fine Tuning

We include a series of notebook+scripts for fine tuning the models.

Gemma 3n

Notebooks

• Gemma 3n Conversational Fine tuning 2B on free Colab T4
• Gemma 3n Conversational Fine tuning 4B with Unsloth on free Colab T4
• Gemma 3n Multimodal Fine tuning 2B/4B with Unsloth on free Colab T4
• Fine tuning Gemma 3n on audio
• Fine tuning Gemma 3n on GUI Grounding
• Fine tuning Gemma3n on video+audio using FineVideo (all modalities)

Scripts

• Fine tuning Gemma 3n on images using TRL
• Fine tuning Gemma 3n on images (script)
• Fine tuning Gemma 3n on audio (script)
• Fine tuning Gemma3n on video+audio using FineVideo (all modalities)

Gemma 3

• Reinforement Learning (GRPO) on Gemma 3 with Unsloth and TRL
• Vision fine tuning Gemma 3 4B with Unsloth
• Conversational fine tuning Gemma 3 4B with Unsloth

RAG

Gemma 3n

• Retrieval-Augmented Generation with Gemma 3n

Before fine-tuning the model, ensure all dependencies are installed:

$ pip install -U -q -r requirements.txt

✨ Bonus: We've also experimented with adding object detection 🔍 capabilities to Gemma 3. You can explore that work in this dedicated repo.