Fine-Tune Gemma 4 Models on Your MacBook

Let’s see how we can fine-tune our first model.

Editor’s Note: We recently published a few deep-dives on Agentic Design Patterns, Workflows and Operations, sourced from the most innovative products and startups of 2026.

You can read the deep dives here, packed with real-world examples, architectures, and workflows you won’t find anywhere else.

Setup on Apple Silicon

Gemma Fine-Tuner targets:

• Python ≥ 3.10, matching its packaging metadata.
• macOS with Apple Silicon for MPS, avoid Rosetta Python for the local/MPS path.
• PyTorch installed separately, because the correct build depends on whether you’re on MPS, CUDA, or CPU-only.

Start by creating a clean virtual environment

Install dependencies

Authenticate with HuggingFace

Install Gemma 4 specific requirements

You can now run the wizard:

Gemma Fine-Tuner ships a 16-row instruction-tuning dataset at data/datasets/sample-text/: translations, summaries, trivia, haiku, and JSON conversion. Small enough to finish in under a minute. Large enough to prove the full pipeline works: data loading, tokenization, LoRA, checkpointing, and export.

Once you press Enter, Gemma Fine-tuner will walk you through the training config with sensible defaults.

Once you confirm the configuration, it will start training.

You can see the training progress in its UI.

That’s it!

Here’s a CLI cheat-sheet that you may find handy.

Streaming and large datasets

Gemma Fine-Tuner also supports a streaming_enabled dataset load path at the dataset utility layer, and the patch/override system explicitly supports streaming iteration, applying overrides and blacklists on the fly during iteration rather than via dataset.map.

Two boundaries to keep straight:

• Text and image fine-tuning are deliberately constrained to non-streaming local CSV in v1, and the dataset loader will reject Granary/BigQuery/GCS streaming adapters for those modalities.
• Audio workflows are where streaming is expected (and where the repo’s release notes emphasize GCS / BigQuery streaming as a differentiator).

From a configuration and code standpoint, the “dataset source adapter” logic is simple and intentionally CSV-centric:

• local CSV non-streaming adapter
• streaming CSV adapter when streaming_enabled is true or source_type indicates streaming/GCS
• and special-handling for dataset “source” naming to align patch directories with streamed dataset identities.

This is trying to keep “what you train on” human-auditable (CSV rows you can inspect), even when the underlying audio assets live remotely.

A practical agentic loop mindset

Now, if you’re building agentic systems on top of a fine-tuned Gemma checkpoint, the high-level loop usually looks like:

• Observation: gather inputs (text + optional audio + optional images)
• Plan: let the model decide whether to call a tool (function calling) or respond directly
• Act: execute tool calls in code, with validation and access control
• Reflect: feed tool outputs back to the model for next-step reasoning
• Commit: produce user-visible output, again schema-checked when needed

Gemma 4 models explicitly position function calling and structured JSON outputs as enabling agentic workflows.

You can use Gemma Fine-Tuner to make the model competent at your domain’s observations and actions, without turning your agent into a prompt spaghetti monster.

A few examples that align closely with the supported modalities and dataset shapes:

• Voice-first internal tooling agent: audio input (calls, meetings), text output (structured JSON summaries, action items, tool calls).
• Screen-understanding support agent: screenshot image input, tool calls to retrieve account state, JSON outputs for ticket creation.
• On-device privacy-preserving assistant: train and run locally with user data constraints, ship merged SafeTensors artifacts for downstream deployment tooling.

If you end up fine-tuning a few models, please share your observations in the comments.