Skip to content

Model Training Guide

This guide covers how to train machine learning models with Arc. Arc handles the complexity of model implementation while you focus on the high-level architecture and training strategy.

Overview

Training a model with Arc involves: 1. Describe your goal - Tell Arc what you want to predict 2. Review Arc-Graph - Arc generates a model architecture specification 3. Train - Arc builds and trains the PyTorch model 4. Monitor - View training progress in TensorBoard 5. Evaluate - Check model performance

Quick Start

The simplest way to train a model:

Build a model to predict diabetes using the processed_diabetes_data table

Arc will: - Analyze your data - Generate an Arc-Graph specification - Train the model - Launch TensorBoard - Report results

Using the /ml model Command

For more control, use the /ml model command:

/ml model --name diabetes_predictor
          --instruction "Build a binary classifier to predict diabetes"
          --data-table processed_diabetes_data

Understanding the Training Process

Arc generates an Arc-Graph specification defining inputs, model architecture, and training configuration. See Arc-Graph Specification for details.

Arc then builds the PyTorch model, loads data, trains for the specified epochs, and tracks metrics.

Monitoring Training

Arc automatically launches TensorBoard at http://localhost:6006 to view training progress, loss curves, and accuracy metrics.

Trained models are saved to ~/.arc/models/<model_name>/ with model weights, Arc-Graph specification, and metadata.

Check Training Jobs

View all training jobs:

/ml jobs list

View specific job details:

/ml jobs status <job_id>

Example

/ml model --name spam_detector
          --instruction "Binary classifier for spam detection with 2 hidden layers (64, 32)"
          --data-table processed_emails

Next Steps