Try Frontier Tuning with a sample tuning environment

In this quickstart, you create a sample tuning environment (otherwise known as a reinforcement learning environment or RLE) in Frontier Tuning and inspect its artifacts. You then evaluate a baseline model, tune the model, and compare the results.

A sample template provides ready-made artifacts and mock data so you can try the workflow before preparing your own business data. The workflow uses the CLI. The API alternative covers only template discovery and environment creation.

Before you begin, confirm service access, the CLI installation, and permission to evaluate and tune models.

Prerequisites

To follow the examples, you need:

  • Access to the tuning service and permission to create an environment.
  • The Frontier Tuning CLI installed. See Install and set up the Frontier Tuning CLI.
  • Permission and available capacity to evaluate and tune a model in your environment.
  • For the API creation alternative, an access token authorized for the service.

What the sample includes

The contents depend on the template you choose. A template can include:

  • Skills: Instructions that describe what the model should do.
  • Evaluation criteria: Requirements for assessing responses.
  • Samples and mock data: Example inputs for exploring the scenario and trying evaluation or training.

An employee-onboarding scenario, for example, could use fictional employee information to generate a checklist. The skill describes the task, and the evaluation criteria check whether the response includes required training and first-week tasks. This example is illustrative; check the template list for available scenarios.

Create a sample tuning environment

Review the available templates and choose a scenario to try. Use either the CLI or the API to create the environment.

Use the CLI

Sign in before discovering templates or creating an environment:

ftune auth login

List the templates:

ftune environments templates list

Replace <template-id> with an ID from the list and review its details:

ftune environments templates get <template-id> -o json

Create a tuning environment from the selected template:

ftune environments create --name "My sample RLE" --template-id <template-id>

Save the environment ID returned by the command. Use it as <environment-id> in subsequent commands.

Use the API

Use the API only for template discovery and environment creation. Then use the CLI to inspect the environment, evaluate models, and run tuning.

Don't include access tokens in shared examples or captured output.

  1. List the available templates. Replace <access-token> with a token authorized for the service.

    GET https://substrate.office.com/KnowledgeGraph/api/v1.0/Environments/Templates
    Authorization: Bearer <access-token>
    
  2. Copy a template ID from the response and retrieve its details. Replace <template-id> with the selected ID.

    GET https://substrate.office.com/KnowledgeGraph/api/v1.0/Environments/Templates/<template-id>
    Authorization: Bearer <access-token>
    
  3. Create an environment with the selected templateId.

    POST https://substrate.office.com/KnowledgeGraph/api/v1.0/Environments
    Authorization: Bearer <access-token>
    Content-Type: application/json
    
    {
      "displayName": "My sample RLE",
      "description": "Explore tuning with ready-made artifacts and mock data.",
      "templateId": "<template-id>"
    }
    
  4. Save the environment ID returned in the response. Use it as <environment-id> below, and sign in to the CLI before continuing:

    ftune auth login
    

Inspect the environment

Review the environment and its contents:

ftune environments get --env-id <environment-id>
ftune skills list --env-id <environment-id>
ftune samples list --env-id <environment-id>
ftune models list --env-id <environment-id>

The skill and sample lists help you understand the task and example inputs. The model list shows models available in the environment.

Before evaluating, confirm that the environment includes Evaluation-type samples. The default evaluation uses these samples. Choose a base model that's available for evaluation and eligible for tuning. Use that model for the baseline evaluation and tuning job.

Evaluate a baseline model

Replace <base-model> with the model you selected and start an evaluation:

ftune evaluate start --env-id <environment-id> --base-model <base-model>

Copy the evaluation job ID from the response. Replace <evaluation-job-id> with that ID to check progress:

ftune evaluate status <evaluation-job-id> --env-id <environment-id>

When the baseline evaluation completes successfully, get the results:

ftune evaluate results <evaluation-job-id> --env-id <environment-id>

Keep the baseline results for comparison with the tuned-model evaluation.

Tune a model

Start a tuning job with a base model eligible for tuning. This example names the resulting model my-sample-model:

ftune tune start --env-id <environment-id> --base-model <base-model> --new-model my-sample-model

Copy the tuning job ID returned by the tuning command. Replace <tuning-job-id> with that ID to check progress:

ftune tune status <tuning-job-id> --env-id <environment-id>

Wait for tuning to complete successfully before evaluating the resulting model.

Evaluate the tuned model

List the models again and identify the tuned model:

ftune models list --env-id <environment-id>

Replace <tuned-model-id> with the resulting model's ID and start a new evaluation:

ftune evaluate start --env-id <environment-id> --base-model <tuned-model-id>

Use the new evaluation job ID, not the baseline evaluation ID, in these commands:

ftune evaluate status <evaluation-job-id> --env-id <environment-id>

After the tuned-model evaluation completes successfully, get its results:

ftune evaluate results <evaluation-job-id> --env-id <environment-id>

Compare the results

Compare the baseline and tuned-model results using the same evaluation samples and evaluation criteria. Look for changes in the criteria that matter to the sample task, rather than assuming tuning improves every response. Review failed tool calls before interpreting differences in scores.

Note

Sample artifacts and mock data are intended for learning and experimentation. Sample results don't establish performance on your business scenarios. Evaluate with representative data before using a tuned model in production.