​
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

1

Navigate to Projects

From the main dashboard, click Projects in the left sidebar, then click + Create Project.GIF showing navigation to Projects section
2

Configure Project Basics

Project Name: Choose a descriptive name that reflects your use caseProject Goals: Add context about your project goalsFollow-up Questions: Answer some questions about your project

​
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.
1

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
Need help with JSONL format? See our Datasets Overview for detailed formatting requirements.

​
Path B: Generate Synthetic Dataset

Generate synthetic dataset from your content (PDFs, web pages, youtube videos)
1

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
2

Configure Advanced Settings (Optional)

Rules & Constraints: Define generation requirements
- Focus on technical accuracy
- Use domain-specific terminology
- Maintain consistent formatting
- Generate questions of varying complexity
Question Format: Guide question structure
Summarize the blog post titles {blog_title} in 5 bullet points.
{blog_post}
Answer Format: Define expected answer style
{bullet_list}
You can also provide QA Examples to better steer the synthetic data generation.Creativity Level: Adjust generation diversity (0-100)
For structured outputs (like JSON extraction), set creativity to 0. For conversational models, use 30-50.
3

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.
Synthetic generation review and confirmation

​
Snapshot Creation

Regardless of which path you chose (upload or synthetic generation), you need to create a snapshot:
1

Overview Dataset

Browse through the QA pairs inside the dataset to ensure quality and relevance.
2

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.
3

Create Snapshot

Create a snapshot of your dataset for fine-tuning:
  • Snapshot Name: Give it a descriptive name
  • Click Create Snapshot to finalize
Give a descriptive name to the snapshot, useful for versioning and easier experiment tracking.

​
Step 3: Fine-tune Your Model

1

Create New Fine-tuning Job

Start by creating a fine-tuning job:
  • Job Name: Give your fine-tuning job a descriptive name
  • Select Model Type: Choose the type of model finetuning you want to apply between reasoning and non reasoning.
2

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
This analysis typically takes 2-5 minutes and helps optimize your training configuration.
3

Configure Experiments

Based on the analysis, the system returns a set of recommended experiments.
4

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
Fine-tuning duration varies from 30 minutes to several hours depending on dataset size, model complexity, and chosen parameters.

​
Step 4: Define Metrics

Before evaluating your model, define the metrics that will measure success for your specific use case:
1

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.
2

Create Custom Metrics

Accept/Update/Delete the previously generated rules to actually create the metric.
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.

​
Step 5: Evaluate Your Model

1

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
2

Start Evaluation

Begin the evaluation process on the validation split, by using the previously defined metrics.
3

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
Evaluation typically takes 5-15 minutes depending on test dataset size. Learn more about evaluation strategies in our Evaluations Guide.

​
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

Deploy Your Model

Learn how to deploy your fine-tuned model for production use.

API Integration

Integrate your model into applications using our API.

Model Experiments

Run experiments to further optimize your model.