Quickstart

Overview

Once you have a clear idea of your AI use case, the next step is building systems that adapt to your specific needs—and fine-tuning your own models is how you get there. At its core, AI comes down to matrix multiplication, captured by the simple equation Ax = y, where A represents your model weights, x your input, and y your output. Prompting only lets you modify x; you have no control over A. Fine-tuning gives you control over A itself, unlocking greater stability and far more precise control over your model's behavior. In short, fine-tuning is what lets you move beyond being another prompt wrapper and build genuinely differentiated AI capabilities.

Fine-tune your first AI model

1. Create Training Project

A training project is a workspace that groups together the fine-tuning jobs you run toward a single goal. For example, if you're building a model to classify support tickets, you might create a project named support-triage and run and manage all related jobs within it.

2. Prepare and Upload Dataset

Before starting a fine-tuning job, you'll need to select a task type and prepare a dataset that conforms to the format described in the Schema Guide.

We support JSON, JSONL, and CSV, though we strongly recommend JSONL for more reliable dataset profiling.

When you upload a dataset, we run a quick profiling pass that validates it against our schema and surfaces key metadata. This helps you confirm the dataset is well-formed and catch outlier samples before training begins.

3. Start your first fine-tune training job

With your dataset uploaded, you're ready to launch your first job. Go to the Training Jobs page and click New Training Job. A job is defined across up to six sections, most of which are prefilled with sensible defaults.

Identity

• Job name (required): A memorable slug that makes each job easy to recognize.
• Description (optional): A short or long note describing the job's purpose.

Base Model

First, choose whether to fine-tune from a foundation model (Pre-trained) or continue from one of your existing checkpoints (Fine-tuned). If this is your first job on the platform, leave Pre-trained selected—you won't have any checkpoints yet.

When selecting a model, the dropdown lists every available base model along with the task types each one supports. Choose any model that supports your target task type.

Technique & task

We currently support SFT and GRPO as training techniques; GRPO is available only for the text-generation task type. Select your target task type here, as it determines your model's structure and output format.

Data

Here you'll see the datasets you uploaded earlier. Select your training dataset, then define your evaluation strategy. You can either split a test set from your training data by specifying a split ratio, or provide a dedicated test dataset (which must also be uploaded under Datasets).

Hyperparameters (Advanced)

The default values reflect our recommended starting points, but you're welcome to adjust them as you experiment. Each hyperparameter includes a tooltip explaining what it does.

Test functions (Optional)

When you've defined an evaluation strategy in the Data section, we query the fine-tuned model at every checkpoint using your test dataset once training completes. Specifying test functions gives you aggregated results per those checkpoints. We recommend reviewing our shared test functions to understand what each one measures.

Click Start training on the right. That's it—your first fine-tuning job is underway.

Check results and deploy your fine-tuned model

1. Track your fine-tuning job

Click the job in the list which you create from the previous step to open its Training Job details page, organized into four sections:

• Overview: Displays the job's metadata (technique, base model, train/test datasets) along with a training loss graph that populates as the job emits loss data points.

  

• Checkpoints: As the job progresses, checkpoints appear here. You can deploy any checkpoint directly from this tab, and if you provided a test dataset, its results are shown here as well.

  

  

• Logs: Real-time logs with a 7-day retention period.

• Configuration: Review the hyperparameters you set for the job.

2. Deploy your fine-tuned artifact

Once a job completes, you can deploy any checkpoint from within the platform. This spins up a dedicated inference server on GPU and provides an endpoint you can integrate into your AI workload.

Open the Checkpoints tab on the training job details page and click Deploy.

Configuring an Inference Engine (deployment) involves four sections, all prefilled—so in most cases you can simply click Deploy to launch your inference server.

Deployment

• Deployment name (required): The name used to reference your model when calling the API.

Target Engine

If you've already created an Inference Engine, you can deploy onto it as long as the base model and task type match. Otherwise, a new Inference Engine will be created to host your artifact.

Engine Configuration

• Engine name: A name for your Inference Engine.
• Autoscaling (optional): Enable autoscaling and define the replica range by setting minimum and maximum replica counts.

Resource Management

We support three ways to manage your Inference Engine and its deployment:

• Keep running: Run the engine continuously until you stop it manually.
• Inactivity timeout: Automatically shut down after a defined period with no requests.
• Schedule: Run only during specific hours and days.

If you're still evaluating the platform, we recommend Inactivity timeout to avoid unnecessary charges.

Click Deploy—and that's it. Your Inference Engine is now live on the platform.

Integrate into your AI workload

From the left sidebar, go to Inference, where you'll find the Inference Engine you deployed in the previous step. Click the engine to open its details page, which is organized into four sections:

• Overview: Lists the deployments attached to this engine. You can add more deployments here as long as they share the same base model and task type, and you can track each deployment's status. Once a deployment shows as Deployed, it's ready to use.

  

• Testing: A built-in UI for testing any deployment that's live. (Image-related task testing isn't supported in the UI yet—coming soon.)

• API: A ready-to-use code snippet for integrating the engine into your service. If you haven't created an API key yet, generate one under Settings > Credentials.

• Monitoring: Collects and displays all of your inference logs.

You can exercise your deployment directly in the Testing UI, or call the API endpoint using the snippet provided in the API tab.

And that's it—you've fine-tuned your own model, deployed your own inference engine, and integrated it into your service.