> ## Documentation Index
> Fetch the complete documentation index at: https://studio-docs.prem.io/llms.txt
> Use this file to discover all available pages before exploring further.

# Get Started with Projects

> Create your first AI development project and learn the complete workflow from dataset creation to model evaluation.

# Creating Your First Project

Projects provide a guided, end-to-end workflow for AI model development. This guide walks you through creating a complete project from start to finish.

## Step 1: General Setup

<Steps>
  <Step title="Navigate to Projects">
    From the main dashboard, click **Projects** in the left sidebar, then click **+ Create Project**.

    <img src="https://static.premai.io/prem-saas-docs/projects/get-started/create-project.gif" alt="GIF showing navigation to Projects section" />
  </Step>

  <Step title="Configure Project Basics">
    **Project Name**: Choose a descriptive name that reflects your use case

    **Project Goals**: Add context about your project goals

    **Follow-up Questions**: Answer some questions about your project
  </Step>
</Steps>

## Step 2: Create and Handle Dataset

You have **two paths** for dataset creation. Choose the one that fits your situation:

### Path A: Upload Existing JSONL Dataset

Upload your existing dataset in JSONL format.

<Steps>
  <Step title="Upload Your JSONL File">
    * **Dataset Name**: Enter a descriptive name
    * **Upload File**: Select your properly formatted JSONL file
    * **Validation**: The system will automatically validate your dataset format

    <Note>
      Need help with JSONL format? See our [Datasets Overview](/datasets/overview) for detailed formatting requirements.
    </Note>
  </Step>
</Steps>

### Path B: Generate Synthetic Dataset

Generate synthetic dataset from your content (PDFs, web pages, youtube videos)

<Steps>
  <Step title="Choose Data Sources">
    Select your input sources for synthetic dataset generation:

    * **Files**: PDF, DOCX, TXT, HTML, PPTX
    * **YouTube Videos**: Individual videos or playlists
    * **Web URLs**: Website content extraction
    * **Mixed Sources**: Combine multiple input types

    **QA Pairs per Source**: Set how many question-answer pairs to generate from each input
  </Step>

  <Step title="Configure Advanced Settings (Optional)">
    **Rules & Constraints**: Define generation requirements

    ```text theme={null}
    - Focus on technical accuracy
    - Use domain-specific terminology
    - Maintain consistent formatting
    - Generate questions of varying complexity
    ```

    **Question Format**: Guide question structure

    ```text theme={null}
    Summarize the blog post titles {blog_title} in 5 bullet points.
    {blog_post}
    ```

    **Answer Format**: Define expected answer style

    ```text theme={null}
    {bullet_list}
    ```

    You can also provide QA Examples to better steer the synthetic data generation.

    **Creativity Level**: Adjust generation diversity (0-100)

    <Tip>
      For structured outputs (like JSON extraction), set creativity to 0. For conversational models, use 30-50.
    </Tip>
  </Step>

  <Step title="Review and Generate">
    Review your configuration summary:

    * Data sources and expected output count
    * Generation settings and estimated cost
      Click **Generate Dataset** to start the process.

    <img src="https://static.premai.io/prem-saas-docs/datasets/get-started/synthetic-gen/step-3-summary.png" alt="Synthetic generation review and confirmation" />
  </Step>
</Steps>

### Snapshot Creation

Regardless of which path you chose (upload or synthetic generation), you need to create a snapshot:

<Steps>
  <Step title="Overview Dataset">
    Browse through the QA pairs inside the dataset to ensure quality and relevance.
  </Step>

  <Step>
    Autosplit the dataset into training and validation split. If you find challenging QA pairs that could be a good test case during evaluation, manually select them and place them in validation set.
  </Step>

  <Step title="Create Snapshot">
    Create a snapshot of your dataset for fine-tuning:

    * **Snapshot Name**: Give it a descriptive name
    * Click **Create Snapshot** to finalize

    <Tip>
      Give a descriptive name to the snapshot, useful for versioning and easier experiment tracking.
    </Tip>
  </Step>
</Steps>

## Step 3: Fine-tune Your Model

<Steps>
  <Step title="Create New Fine-tuning Job">
    Start by creating a fine-tuning job:

    * **Job Name**: Give your fine-tuning job a descriptive name
  </Step>

  <Step title="Dataset Analysis">
    The system will analyze your dataset automatically:

    * **Data Quality Assessment**: Checks for formatting and consistency
    * **Content Analysis**: Analyzes patterns and complexity
    * **Recommendations**: Suggests optimal training models and parameters

    <Note>
      This analysis typically takes 2-5 minutes and helps optimize your training configuration.
    </Note>
  </Step>

  <Step title="Configure Experiments">
    Based on the analysis, the system returns a set of recommended experiments.
  </Step>

  <Step title="Run Experiments">
    Start the fine-tuning process:

    * **Experiment Selection**: Choose which experiments to run
    * **Monitor Progress**: Track training progress in real-time
    * **Compare Results**: View performance across different configurations

    <Note>
      Fine-tuning duration varies from 30 minutes to several hours depending on dataset size, model complexity, and chosen parameters.
    </Note>
  </Step>
</Steps>

## Step 4: Define Metrics

Before evaluating your model, define the metrics that will measure success for your specific use case:

<Steps>
  <Step title="Generate Rules">
    Describe your evaluation needs and let our AI generate the rules for you.
    Provide:

    * **metric name**
    * **metric description**

    Our AI will generate a set of **should follow** and **should not follow** rules.
  </Step>

  <Step title="Create Custom Metrics">
    Accept/Update/Delete the previously generated rules to actually create the metric.
  </Step>
</Steps>

<Note>
  Well-defined metrics are crucial for meaningful evaluation. Take time to think about what "success" looks like for your specific use case. Learn more in our [Evaluation Metrics Guide](/evaluations/metrics).
</Note>

## Step 5: Evaluate Your Model

<Steps>
  <Step title="Choose Metrics">
    Select the metrics you want to use for evaluation:

    * **Your Custom Metrics**: The metrics you defined in the previous step
    * **Available Metrics**: Pre-built metrics from the Prem library
  </Step>

  <Step title="Start Evaluation">
    Begin the evaluation process on the validation split, by using the previously defined metrics.
  </Step>

  <Step title="Review Evaluation Results">
    Analyze the results to understand model performance:

    * **Metric Scores**: See how each model performed on each metric
    * **Detailed Breakdowns**: Examine individual response quality
    * **Identify Weaknesses**: Find areas where models need improvement

    <Note>
      Evaluation typically takes 5-15 minutes depending on test dataset size. Learn more about evaluation strategies in our [Evaluations Guide](/evaluations/get-started).
    </Note>
  </Step>
</Steps>

## Project Complete

Congratulations! Your project is now complete. You have:

* ✅ **Created or uploaded a quality dataset**
* ✅ **Successfully fine-tuned your model**
* ✅ **Defined custom metrics for your use case**
* ✅ **Validated performance through evaluation**
* ✅ **Generated a production-ready AI model**

### Next Steps

<CardGroup cols={2}>
  <Card title="Deploy Your Model" icon="rocket" href="/inference/self-host">
    Learn how to deploy your fine-tuned model for production use.
  </Card>

  <Card title="API Integration" icon="code" href="/api-reference/introduction">
    Integrate your model into applications using our API.
  </Card>
</CardGroup>

<CardGroup cols={1}>
  <Card title="Model Experiments" icon="flask" href="/finetuning/experiments">
    Run experiments to further optimize your model.
  </Card>
</CardGroup>