Agent Platform Model Tuning
A guided end-to-end workflow for fine-tuning open or Gemini models on Google Agent Platform.
The evidence shows confirmations for project and region, dataset columns, model, hyperparameters, job submission, and deployment, plus stop conditions for inaccessible data or GCS objects. However, least-privilege IAM roles, sensitive-data handling, data-flow disclosure, deployment rollback, and an integrated cancellation/recovery path are incomplete. Authentication, API enablement, dependency installation, uploads, and deployment create external effects, so points are deducted.
The skill provides a staged workflow, input checks, dataset validation, a fixed split seed, and some actionable error feedback. However, referenced requirements.txt, tune_open_model.py, and tune_gemini_model.py files are absent from the supplied material; messages_gemini is inconsistent with the validator's supported formats; and a gcloud configuration command appears nonstandard. Key paths cannot be reproduced from a static read, so the score remains below the static 10-point ceiling.
The intended users, Agent Platform tuning scenarios, and exclusions for non-Agent-Platform training, deployment, and endpoint management are reasonably clear. Model-category and dataset-state decision points are also specified. Trigger precision, non-fit ranges, Chinese-language support, and mainland-China reachability are underdocumented, while core dependencies include Google Cloud, Hugging Face, and overseas documentation services, so points are deducted.
The documentation uses an overview, decision tree, phased workflow, references, installation notes, commands, limitation warnings, Apache-2.0 licensing, and support/contribution paths. It lacks explicit versioning, a changelog, named maintenance responsibility, and a clear update path. Referenced dependencies and scripts are incomplete, and format terminology is inconsistent, so points are deducted.
The skill theoretically covers environment setup, dataset conversion and validation, upload, configuration, submission, monitoring, and deployment. However, key submission scripts and the requirements file are not supplied, the Gemini path may require manual work, the cost script only handles messages-shaped data, and result validation is limited. The static ceiling is therefore 7, with additional deductions for incomplete direct usability.
The supplied Markdown, Python scripts, model table, and official documentation links provide limited auditability. There are no committed tests, CI coverage, third-party execution records, or corroborating evidence types, and several capability and currency claims cannot be independently confirmed from the supplied files. The score is therefore low.
- Do not perform authentication, API enablement, uploads, or deployment before verifying least-privilege IAM, data sensitivity, and the target project.
- Confirm that requirements.txt, tuning scripts, and Gemini data-format support actually exist, and resolve inconsistencies between the scripts and documentation.
- Core operation depends on Google Cloud, Hugging Face, and external documentation services; mainland-China reachability and enterprise proxy requirements need separate verification.
- Add explicit rollback, cancellation, and failure-recovery procedures before deployment.
What does this skill do, and when should you use it?
This skill is for fine-tuning large language models with Google Agent Platform, covering both open models and Gemini models. It starts with authentication, API, IAM, and virtual-environment setup, then handles dataset preparation, JSONL validation, GCS upload, model and hyperparameter configuration, job submission, and monitoring. It also provides deployment guidance after tuning completes. It is not intended for training outside Agent Platform, endpoint deployment, or endpoint management.
Checks and configures the gcloud CLI, project, and region; guides API enablement; and verifies relevant service agents and user permissions. It uses Python scripts to prepare, convert, and validate JSONL data, asks whether to create an 80/20 validation split when appropriate, and uploads datasets to Google Cloud Storage. It recommends models and hyperparameters from the referenced catalog and tuning guide, calculates a rough Open Model cost estimate, and submits tuning jobs only after user confirmation. It also guides job monitoring and deployment of a successfully tuned model.
- A machine-learning engineer wants to fine-tune an open model on Google Cloud and needs dataset validation, cost estimation, and job-submission guidance.
- A team wants to tune a Gemini model with Agent Platform and needs help preparing the environment, data format, and submission path.
- A developer has training data that is not yet valid JSONL and needs its message roles and schema checked.
- A Google Cloud user needs to upload prepared training data to GCS and run a tuning job.
- A user has a completed tuning job and needs guidance deploying the resulting model on Agent Platform.
What are this skill's strengths and limitations?
- Covers the tuning lifecycle from environment setup through data preparation, configuration, submission, monitoring, and deployment.
- Supports both open models and Gemini models with separate data-format guidance.
- Includes explicit checks for dataset validity, column mapping, GCS accessibility, and user approval.
- Provides a rough cost-calculation workflow and model-recommendation process for open models.
- Requires gcloud, a Python virtual environment, Google Cloud APIs, IAM permissions, and GCS; it is not a standalone tuning workflow.
- Model recommendations depend on referenced files such as references/models.md and references/tuning_guide.md, which were not included here, so their specific catalog contents cannot be verified.
- The source provides no test suite, performance measurements, or evidence of execution on specific platforms.
- The Gemini submission path requires a runtime check for the script; if it is absent, the user must submit through the console or SDK.
- It does not cover training outside Agent Platform, endpoint deployment, or endpoint management.
How do you install this skill?
This skill is distributed as part of the google/skills collection. Run: npx skills add google/skills. Select Agent Platform Model Tuning during installation; the source does not document a standalone installation command for this skill.
How do you use this skill?
In an Agent Skills-compatible client, make the model category explicit, for example: “I want to fine-tune an Open Model for customer-support responses using Agent Platform.” You should have a Google Cloud project, region, dataset, and GCS access available. The workflow requests confirmation of the project and region, column mapping, model and hyperparameters, job-submission command, and deployment command at the relevant stages.
How does this skill compare with similar options?
Its scope is distinct from the collection’s Agent Platform Model Garden Deploy Skill and Agent Platform Endpoint Management Skill: this skill handles model tuning, while those skills address endpoint deployment and endpoint management.