General Questions

What is Prem?

Prem is a AI platform that enables you to create, fine-tune, evaluate, and deploy custom AI models. It’s designed to make AI integration seamless and efficient, allowing you to focus on developing your product without dealing with AI complexity.

What can I do with Prem?

With Prem, you can:
  • Access best-in-class AI models without vendor lock-in
  • Autonomously fine-tune models with your own datasets
  • Evaluate models using built-in metrics and tools
  • Test models with the built-in Playground
  • Monitor your models’ performance over time with Stats

How do I get started with Prem?

The best way to start is by creating your first Project, which guides you through the complete AI development workflow from dataset creation to model evaluation. Alternatively, you can follow our Quick Start Guide for a direct approach.

Does Prem require machine learning expertise?

No, Prem is designed to be accessible to users without machine learning expertise. Our Autonomous Fine-Tuning feature handles the complexities of model training for you.

What are Projects and why should I use them?

Projects provide a guided, end-to-end workflow that orchestrates your entire AI development process. They’re especially useful for beginners as they connect dataset creation, fine-tuning, and evaluation into a single, streamlined experience. Learn more in our Projects Overview.

Projects

Do I need to prepare data before creating a Project?

No! Projects support two approaches: you can upload existing JSONL datasets if you have them, or generate synthetic datasets directly from various sources like PDFs, websites, and documents. This flexibility makes Projects perfect for any starting point.

What types of files can I use to generate synthetic datasets?

Projects can generate datasets from PDF, DOCX, TXT, HTML, PPTX files, YouTube videos, website URLs, or any combination of these sources. The platform automatically extracts content and creates question-answer pairs for training.

How long does a complete Project take?

A typical Project takes 1-3 hours total: dataset generation (10-30 minutes), fine-tuning (30 minutes to 2 hours), and evaluation (5-15 minutes). The exact time depends on your dataset size and model complexity.

Datasets

What format should my dataset be in?

Datasets should be in JSONL (JSON Lines) format, where each line represents a single conversation example. Each example should contain a “messages” field with an array of message objects that have “role” and “content” fields.

What roles are included in a dataset?

Each conversation in your dataset should include three roles:
  • “system”: Provides context and instructions for the model’s behavior
  • “user”: Represents what a human user would say or ask
  • “assistant”: Contains the ideal response you want the model to learn

How do I create a dataset in Prem?

You have two main options: upload an existing JSONL file or generate a synthetic dataset from various sources (PDFs, websites, videos). The easiest way is through Projects, or you can create standalone datasets following our Get Started with Datasets guide.

Can I create datasets without existing training data?

Yes! Prem’s synthetic data generation can create high-quality datasets from raw content like PDFs, documents, websites, and videos. This is perfect when you have domain knowledge but no formatted training data. See our Synthetic Data Guide for details.

Can I enrich my existing dataset?

Yes, Prem offers dataset enrichment to add synthetic data to existing datasets, helping you expand coverage and improve model performance. Learn more in our Enrich Dataset documentation.

How much data do I need for a good dataset?

Quality matters more than quantity. For most use cases, 100-500 high-quality examples work well, though complex domains may need 1000+ examples. Projects help you generate the right amount of data for your specific use case.

What is a dataset snapshot?

A snapshot is a fixed version of your dataset that can be used for fine-tuning. Creating snapshots allows you to preserve specific versions of your dataset for training or comparison purposes. Learn more in the Create Snapshot guide.

Fine-Tuning

What is Autonomous Fine-Tuning?

Autonomous Fine-Tuning is Prem’s approach to creating custom models without requiring machine learning expertise. It automatically handles the complexities of the fine-tuning process for you.

Why should I fine-tune a model?

Fine-tuning offers several benefits:
  • Adapts models to specific tasks or domains
  • Improves performance for your specific use cases
  • Requires less data and resources than training from scratch
  • Allows customization to your unique requirements
  • Helps overcome limitations in the original model

How does the fine-tuning process work?

The basic process involves:
  1. Creating a snapshot of your dataset
  2. Fine-tuning your selected model on this dataset snapshot
  3. Evaluating the resulting model to verify performance
For the complete guided experience, use Projects which automates this entire workflow.

How long does fine-tuning take?

Fine-tuning duration depends on your dataset size and the model you’re using. Typically, the process can take anywhere from 30 minutes to a few hours. Projects provide real-time progress tracking so you know exactly when your model will be ready.

Evaluations

What are model evaluations?

Model evaluations are systematic assessments of how well AI models perform on specific tasks, measuring aspects like accuracy, reliability, fairness, safety, and alignment with human preferences.

Why are evaluations necessary?

Evaluations are critical for:
  • Quality assurance before deployment
  • Identifying areas needing improvement
  • Optimizing costs by selecting the most efficient model
  • Mitigating risks by detecting potential biases or unsafe outputs
  • Tracking performance progress over time

How does Prem’s evaluation system work?

Prem offers an integrated evaluation system that:
  • Seamlessly evaluates models within the same platform
  • Uses your actual datasets for real-world performance testing
  • Provides comparative analysis between different models
  • Enables continuous monitoring of model performance

What metrics are used for evaluation?

Prem offers various evaluation metrics depending on your use case. For specific information about available metrics, please refer to our Evaluations Overview documentation.

Models

What models are available on Prem?

Prem offers a wide range of models from providers like Anthropic (Claude), OpenAI (GPT), Meta (Llama), Google (Gemma), and others. For a complete and up-to-date list, see our Available Models page.

Can I use my own custom models?

Yes, Prem allows you to fine-tune existing models to create your own custom versions tailored to your specific use cases.

How do I choose the right model for my use case?

The best model depends on your specific requirements, including:
  • Task complexity
  • Performance needs
  • Resource constraints
  • Budget considerations You can use our Playground and Evaluation tools to compare different models for your specific use case.

Pricing and Support

How is Prem priced?

For the most current pricing information, please contact our sales team at [email protected].

Where can I get help if I have issues?

You can get support by:

Do you offer enterprise support?

Yes, we offer enterprise support options. Please contact us at [email protected] to discuss your specific needs.

Where can I learn about the latest news from Prem?

You can stay updated on the latest features and improvements by checking our Release Notes or following our Blog.