Agent Platform Tuning skill: fine-tune Gemma, Llama, Qwen or Gemini on Google Cloud
Walks your AI assistant through fine-tuning an open model or a Gemini model on Google Cloud's Agent Platform.
- Installs
- 9K
- on skills.sh
- Stars
- 21.8K
- on GitHub
- Files
- 11
- besides SKILL.md
- Updated
- Nov 2026
- last repo commit
agent-platform-tuning has your assistant check your Google Cloud project, prepare and validate a training dataset, upload it to Cloud Storage, pick a model from a catalog, and submit a tuning job. It estimates the cost of open model jobs with a bundled script and asks for a Yes or No before it submits a job or deploys the result. It needs gcloud, Python and a Google Cloud project, and tuning and deployment run on your project. It lives in Google's own google/skills repo.
Good for
- ML engineers who want to tune Gemma, Qwen, Llama or Gemini models on Google Cloud
- Teams with a CSV, JSON, Parquet or JSONL dataset to turn into a tuning job
- Developers who want a cost estimate and a yes/no check before a job is submitted
Install agent-platform-tuning
Claude Code
- 1
Download the skill folder from https://github.com/google/skills/tree/HEAD/skills/cloud/agent-platform-tuning.
- 2
Put it at ~/.claude/skills/agent-platform-tuning/ so that SKILL.md sits directly inside.
- 3
Start a new session; the skill loads when its description matches the task.
Requires gcloud CLI, Python 3.
Sign in first with gcloud auth login and have a Google Cloud project and Cloud Storage bucket. SKILL.md checks python3 -c "import vertexai, google.genai, google.cloud.storage, datasets" and runs pip install -r references/requirements.txt only if that fails.
Built from the package data we hold. Not yet tested by us.
What's in the skill
When it loads
Agent Platform Model Tuning. Use when you need to fine-tune open models or Gemini models using Agent Platform infrastructure. Don't use for model training outside Agent Platform, model deployment to endpoints (use `agent-platform-deploy`), or managing serving endpoints (use `agent-platform-endpoint-management`).
Sections in SKILL.md
- Overview
- Workflow Decision Tree
- Phase 0: Environment & IAM Setup {#phase-0}
- 0.1 Authentication & Project Context
- 0.2 Location
- 0.3 Enable APIs
- 0.4 IAM Permissions
- 0.5 Python Dependencies
- Phase 1: Dataset Preparation & Upload {#phase-1}
- 1.0 Dataset Discovery & Confirmation
- 1.1 Formatting & Validation
- 1.2 Upload
- Phase 2: Model Configuration & Recommendation {#phase-2}
- 2.1 Configuration
- 2.2 Calculating Cost (Open Models Only)
- Phase 3: Tuning Job Execution {#phase-3-tuning-job-execution}
Key facts
- Repo
- google/skills
- License
- Apache-2.0
- Installs on skills.sh
- 9,027 as of 2026-11-02
- Last repo commit
- 2026-11-01
- What is inside
- 12 files including SKILL.md: 6 Python scripts and 5 reference files (4 markdown, 1 requirements.txt)
- Skill version
- 1.0.1 (frontmatter)
What it includes
Asks for project, region and model category before acting
Nine checks: project and region, open or Gemini model, environment, dataset, column mapping, configuration, job, completion and deployment. It stops and asks if the project or region is missing.
Try: “Fine-tune an open model on my support-tickets.csv in project my-proj, region global.”
Read ·
Decision treegcloud auth, location, APIs, roles
Checks gcloud and auth, picks a location (global is recommended for open models), runs gcloud services enable for the Agent Platform and Cloud Storage APIs, and checks roles for you and the service agents.
Runs code ·
Phase 0: Environment and IAMConverts CSV, JSON or Parquet to JSONL and validates it
Formats data for tuning, can split off a 10 percent validation set, and has a validate-only mode. SKILL.md says to ask before splitting.
Write ·
scripts/prepare_dataset.pyCopies the dataset to your bucket
Uploads the JSONL to a timestamped folder in the bucket you name, used exactly as written.
Runs code ·
Upload with gcloud storage cpSuggests public datasets for a task
references/hf_datasets.md lists datasets and tells the assistant to load them with the datasets library, ask you to confirm the split and columns, and stream large ones.
Network ·
Hugging Face datasetsLists live Model Garden IDs
Lists up to 20 matching model IDs through the vertexai SDK so the assistant picks from real IDs.
Network ·
scripts/list_models.pyEstimates open-model tuning cost
Reads your dataset (including from Cloud Storage) and a model entry to print an estimated dollar figure that goes into the confirmation card. SKILL.md says it can differ from actual billing.
Network ·
scripts/calculate_cost.pySubmits an open-model tuning job
Needs an --output_uri bucket path. Gemini jobs go through the google.genai or vertexai SDK in Python snippets instead.
Runs code ·
scripts/tune_open_model.pyPolls job status in the background
Runs only if you agree. Otherwise you check the Cloud Console link yourself.
Network ·
scripts/monitor_tuning_job.pyCancels a tuning job
Takes project, location and job ID and sends a cancel request. SKILL.md does not mention it.
Runs code ·
scripts/cancel_tuning_job.pyRegisters the tuned model and deploys it to a new endpoint
Runs gcloud ai commands to upload the model, create an endpoint and deploy to it, after you confirm the commands. It needs a real region, not global.
Runs code ·
Phase 5: Deploy
From the official docs
Permissions
High risk
The skill runs gcloud and Python against your Google Cloud project: it enables APIs, uploads data, starts paid tuning jobs and can deploy an endpoint. Use a test project and read each confirmation card before you say Yes.
What it can do
- Read your data
- Create and change things
- Delete things
- Read-only mode
Watch out for
Tuning jobs and endpoints cost money
Jobs run on the Google Cloud project you name; the charge lands there (our reading). SKILL.md gives a dollar estimate for open-model jobs only (scripts/calculate_cost.py) and says it can differ from actual billing; it names no cost script for Gemini jobs. Check the figure on the confirmation card before you approve. github.com
Phase 5 creates a dedicated endpoint
After a job succeeds, the deploy script registers the tuned model, creates an endpoint and deploys it with the machine type and accelerator from list-deployment-config, after a Yes/No card. It writes flags files to /tmp. SKILL.md gives no undeploy or delete step for the endpoint, so you must remove it yourself; see the cleanup steps in the agent-platform-deploy skill. github.com
It runs gcloud with --quiet and changes your project
Phase 0 runs `gcloud components update --quiet`, and gcloud services enable changes the project. Job submission and deployment ask for a Yes/No confirmation first, per the SKILL.md. github.com
Your training data goes to Cloud Storage and Google's tuning service
The dataset is uploaded to the bucket you name and used as the training dataset. Do not use data you cannot send to Google Cloud. github.com
It may install Python packages
If the import check fails it runs pip install -r references/requirements.txt, which pins google-cloud-aiplatform, datasets, google-cloud-storage and google-genai. github.com
A cancel script is bundled
scripts/cancel_tuning_job.py sends a cancel request for a job ID. SKILL.md does not tell the assistant to call it, but it sits in the folder. github.com
Technical details
| Turning on read-only | No read-only mode. The environment check is read-only; later phases write. |
|---|---|
| Talks to | aiplatform.googleapis.com and storage.googleapis.com (Google Cloud). console.cloud.google.com (job links). Hugging Face (dataset downloads through the datasets library) |
| Sign-in and keys | Uses your gcloud sign-in (gcloud auth login) and the Python SDKs' credentials. |
| Scopes |
|
Questions
What does agent-platform-tuning do?
It guides your assistant through the whole tuning workflow on Google Cloud: check the project, prepare a JSONL dataset, upload it, choose a model, estimate cost, submit the job, monitor it and deploy the result.
How do I install it?
Use the install box at the top for Claude Code, Codex or the skills CLI. You also need the gcloud CLI signed in, Python 3, and a Google Cloud project. The box has no Claude app upload because the skill runs gcloud and Python on your machine.
Is this an official Google skill?
It is in Google's own google/skills repo, which describes itself as agent skills for Google products. The repo README does not name an individual author for this skill.
Which models can it tune?
Open models from a catalog in references/models.md (examples in SKILL.md are Gemma 3, Qwen 3 and Llama 3.1) and Gemini models. For Gemini it recommends only gemini-2.5-flash or gemini-2.5-pro.
Will it spend money without asking?
SKILL.md requires a Yes/No confirmation card before a job is submitted and before deployment, and it shows the estimated cost for open models. Read the card, because the estimate can differ from actual billing.
Does it need an API key?
No key is named. It uses your gcloud sign-in. Hugging Face datasets are loaded through the datasets library.
Is it safe?
It runs commands that change a billed Google Cloud project: enabling APIs, uploading data, starting tuning jobs and deploying endpoints. Use a test project first and read every confirmation. See the permissions section.
What about deploying and managing endpoints?
The description says to use agent-platform-deploy for deploying and agent-platform-endpoint-management for serving endpoints, yet Phase 5 also has its own gcloud deployment steps.
Information
updated Nov 2, 2026
- Repository
- google/skills
- Folder
skills/cloud/agent-platform-tuning
Sources
Docs, issue trackers and discussions this page is built from.
Show all 5Hide
- agent-platform-tuning SKILL.md · github.com
- requirements.txt · github.com
- references/hf_datasets.md · github.com
- google/skills README · github.com
- prepare_dataset.py · github.com