Fine-tune for Less than $1 on AI Studio: A Quick Tutorial

Think fine-tuning is complicated or expensive? On AI Studio, you can fine-tune a base model for less than $1* and have a custom model ready to deploy in just a few minutes. By uploading your own dataset and following a few simple steps, you can create a model that delivers outputs for your specific needs. No matter if you want it for improving relevance, domain-specific responses or testing new use cases, AI Studio makes fine-tuning fast and accessible, even for first-time users.

In our quick tutorial below, we fine-tune a base model with a domain-specific dataset to build a custom model that delivers more relevant responses in Playground testing.

How to Fine-Tune a Base Model on AI Studio

Follow the steps below to get started with fine-tuning on AI Studio:

Step 1: Sign in and Access AI Studio

1. Go to the AI Studio here.
2. Sign in using your Hyperstack account. If you are new to Hyperstack, check out our getting started guide here.

Once logged in, you will be able to view the dashboard below and start fine-tuning with the following steps:

Step 2: Upload Your Training Data

To begin fine-tuning, you need a dataset. In this tutorial, we’re using a sample dataset containing 15 prompt-response pairs in .jsonl format. 

Each entry simulates a short interaction where the user asks a question about Hyperstack or AI Studio and the assistant provides a relevant answer. 

This dataset helps train the model to act as a knowledgeable assistant on Hyperstack’s terminology, features and services. The base model doesn’t have these capabilities before fine-tuning.

Follow these steps to upload your dataset:

1. Open the Data Page

In Hyperstack AI Studio, go to the Logs & Datasets page.

2. Upload Your .jsonl File

Click here to download the example dataset for fine-tuning.

Once downloaded, click the Upload Logs button in the top-right corner. Then select the .jsonl file or drag and drop it into the upload area.

3. Add Tags and Upload

- Enter at least one tag to help categorise your logs (e.g., customer-service).
- Click on Validate and Upload. The system will automatically check your file format and  structure and upload the logs if validation succeeds.

Step 3: Start a Fine-Tuning Job

1. Go to My Models and click on New Fine-Tuning Job.

2. Select the base model you want:

• Mistral Small 24B Instruct 2501
• Llama 3.3 70B Instruct
• Llama 3.1 8B Instruct

3. Give your fine-tuning job a unique name. Here we name it: test-model-123

Step 4: Configure Your Settings

Once your dataset is uploaded, it’s time to configure your fine-tuning job with the following steps.

1. Select Training Data: Choose the source of logs for training:

• All Logs: Use all logs in your account.
• By Tags: Select logs with specific tags.
• By Dataset: Select logs from an existing dataset.
• Upload Logs: Upload new logs for this training run. You can choose to save these logs with custom tags or not save them at all. Make sure your logs follow the JSONL File Format guidelines.

2. Context and Sequence Length Limit: Each training example (prompt + expected response) must fit within a sequence length limit of 2048 tokens. After deployment, your fine-tuned model can handle up to 8192 tokens for inference conversations.

3. Adjust Advanced Settings (Optional): Click Advanced Settings to customise training hyperparameters, such as:

• Epochs
• Learning rate
• Batch size and more

If no changes are made, default values will be applied. 

Step 5: Train Your Model

Review your configuration and click Create to start fine-tuning. The following steps will begin automatically:

• Validation: Logs are checked and filtered. Any invalid entries will be skipped.
• Training Begins: Resources are allocated and the job starts.
• Training In Progress: The job runs in the background while you monitor progress.

Want to stop fine-tuning? Click on “Cancel Training” anytime.

Step 6: Review Training Results

1. Click View Model Details for training status, loss over time and cost per million tokens.

2. Go to Model Details and click Training Metrics for detailed metrics.

3. Check key insights like Training loss, Validation loss and Performance comparison (start vs. end)

4. The Training Metrics below indicate clear improvement in model performance as a result of our fine-tuning. The model is learning and generalising better on unseen data.

Training loss dropped from 2.577 → 0.361, validation loss from 2.641 → 1.616.

Step 7: Deploy the Fine-Tuned Model

1. Once training finishes, the model status changes to Ready to Deploy.

2. Click the Status toggle to start deployment. Status will briefly show Deploying.

3. When complete, it updates to Deployed and your model is now live and ready to use.

Step 8: Testing The Model in The Playground

1. Go to the Playground page in AI Studio.

2. From the Model dropdown, select your fine-tuned model (e.g., test-model-123).

3. Adjust Advanced Parameters to customise generation behaviour. (Optional) 

4. Enter a query in the text box → press Enter to see the response.

Here we ask the fine-tuned model: What is Hyperstack AI Studio?

Context and Limits

The training examples must fit within 2048 tokens. Fine-tuned models on AI Studio support up to 8192 tokens for inference after deployment.

Comparison with Base Model

To compare the fine-tuned model with the base model, click Compare Side-by-Side to test it against the base model.

• Select both models from the dropdowns.
• Enter your prompt and compare outputs in real time. We put: What is Hyperstack AI Studio?
• The fine-tuned model answers Hyperstack-specific questions clearly, while the base model Mistral Small 24B Instruct 2501 cannot.

Playground via API

You can enable API Mode in the AI Studio Playground to generate a cURL command for integrating the completions API into your business application. Simply toggle the API option in the Playground interface to view the command.

For complete information on using the API, refer to the Model Inference API documentation.

*Disclaimer: Finetuning for under $1 applies only to the example dataset in the tutorial for Llama 3.1 8B and Mistral Small 24B using default hyperparameters. Actual charges may vary based on workload or dataset size.

FAQs

What is AI Studio?

AI Studio is a full-stack Gen AI platform built on Hyperstack’s high-performance infrastructure. It is a unified platform that helps you go from dataset to deployed model in one place, faster.

Which models can I fine-tune on AI Studio?

You can fine-tune these popular open-source models:

• Llama 3.3 70B
• Llama 3.1 8B Instruct
• Mistral Small 24B Instruct

Can I fine-tune on AI Studio for under $1?

You can fine-tune for under $1 only when using the example dataset in this tutorial with Llama 3.1 8B or Mistral Small 24B and default hyperparameters. Costs may increase with larger datasets, different models, or custom training settings.

Can I test my fine-tuned model outputs before deployment on AI Studio?

Yes. The built-in Playground allows you to interact with your fine-tuned model in real time, test outputs and compare them against the base model before deploying.

What dataset was used in this tutorial?

This tutorial uses a sample .jsonl dataset with 15 prompt-response pairs. Each entry simulates a short interaction about Hyperstack or AI Studio, helping the fine-tuned model understand domain-specific terminology and provide accurate responses.

How do I upload my dataset to Hyperstack AI Studio?

It is simple. Just got to Logs & Datasets, click Upload Logs, select or drag your .jsonl file, add at least one tag and click Validate & Upload. The system checks your file and prepares it for training.

Can I try Hyperstack AI Studio without my own dataset?

Yes! You can start immediately using our sample dataset (Click here to download the example dataset for fine-tuning), which contains 15 prompt-response pairs designed to help you test fine-tuning and see results quickly.

How do I deploy a fine-tuned model on Hyperstack AI Studio?

After training is complete, the model status will show Ready to Deploy. Click the Status toggle to deploy. The status will briefly show Deploying, then update to Deployed, making your model live and ready to use.