Fine-Tuning

Fine-tuning allows you to continue training an AI model on your use-case-specific data. ModelStudio uses LoRA (Low Rank Adapter) methodology for efficient fine-tuning.

What is LoRA?

LoRA freezes the pre-trained weights of the model and instead introduces trainable rank decomposition matrices into the layers. This reduces the number of trainable parameters while maintaining quality comparable to full fine-tuning.

Configuration Parameters

Model Configuration

A descriptive name for your fine-tuned model that will appear in selection menus and help identify its purpose or version.

Training Dataset

The full dataset is used for training and needs to be created beforehand. See Datasets for creation methods.

Base Model

The pre-trained foundation model to fine-tune. Select based on your use case:

Text tasks: Qwen3 0.6B, Qwen3 1.7B, or Granite 3.3 2B
Vision tasks: Qwen2.5-VL 3B

See Base Models for detailed information.

Training Settings

Training parameters can be configured manually or automatically optimized.

Recommendation: Use auto-optimization unless you have specific parameter requirements.

Manual Configuration

Learning Rate (default: 2e-4)

Controls how quickly the model adapts to your data during training
Higher learning rates enable faster learning but can lead to unstable training
Lower learning rates result in slower, more stable adaptation

Batch Size (default: 16)

Number of training examples processed simultaneously
Larger batch sizes provide more stable gradients but require more memory
Smaller batch sizes use less memory but may have noisier gradients

Epochs (default: 3)

Number of complete passes through the training dataset
Typically 3-10 epochs for fine-tuning
Too many epochs can lead to overfitting

Overfitting: Training for too many epochs may cause the model to memorize training data rather than learning generalizable patterns. Monitor validation metrics.

Fine-Tuning Jobs

Job Status

Fine-tuning jobs appear in the job overview with their current status:

Queued
Running
Completed
Failed

Job Details

Each job provides detailed information:

Training step-by-step status
Dataset fingerprint (unique identifier)
Worker assignment
Training timeline
Parameter values used

Testing Your Model

Playground

After training completes, test your model in the interactive playground:

Test up to 16 input examples simultaneously
10MB total size limit across all inputs
Upload images or PDFs for vision models
Automatic OCR for PDF text extraction (playground only)

Note: The playground includes OCR and document processing pipelines that are not available via API. Implement equivalent preprocessing in your production application.

Use at least 50-100 training examples
Ensure diverse, representative data
Maintain consistent formatting
Include edge cases

Parameter Selection

Start with auto-optimization
Monitor training metrics
Adjust only if needed
Avoid over-training

Next Steps

Test in Playground - Validate model performance
API Integration - Deploy to production
Iterate - Add data and retrain to improve