Quick Start

Get from your documents to a fine-tuned model in a few simple steps.

1. Create a Project

After signing in at app.altai.dev:

1. From the navigation bar, expand the project selector, and select Create new project.
2. Enter a project name and submit. Your project will be bound to your current organization.
3. Your new project will be selected automatically, giving you access to your Runs and Files.

2. Generate a Synthetic Dataset

Turn your files into a conversational dataset. From the dashboard:

1. Click New run and choose Synthetic dataset.
2. Step 1 — Name & tasks: Set a run name and choose the task types you want reflected in the dataset (e.g., summarization, question answering, RAG-style behavior).
3. Step 2 — Assistant prompt: Provide a prompt that guides how the responder side of the dialogs should behave. You can use the Improve prompt button or save it to your Project Prompt Library for later use.
4. Step 3 — Choose files: Select or upload project files (PDF, DOCX, RTF, HTML, TXT) as sources.
5. Step 4 — Choose when to stop: Configure generation stop conditions such as a fixed number of samples, context coverage, token count, or a credit budget.
6. Review and start the run. Once generation completes successfully, you can preview the paginated dataset.

3. Fine-tune a Model

Train a base model using your newly generated dataset:

1. Click New run and choose Fine-tune.
2. Step 1 — Name & datasets: Provide a run name and select your dataset.
3. Step 2 — Tasks & prompt: Pick at least one task and provide an assistant prompt (you can select the one saved in your Project Prompt Library).
4. Step 3 — Model: Choose a base model from the list offered (e.g., Qwen3, Gemma, Llama variants).
5. Step 4 — Cost & run: Review any hyperparameters (such as learning rate and LoRA settings) and confirm to start training.

4. Download Results

When training completes, navigate to the Run detail page for your fine-tuning run. From the Results tab, you can view the run results, read the logs, and download your fine-tuned model artifacts.