Model Tuning
Customize and fine-tune machine learning models for your specific use cases
Model tuning allows you to adapt Mixpeek’s underlying machine learning models to your specific domain, improving performance on specialized tasks and content types.
Overview
Model tuning in Mixpeek lets you customize and fine-tune feature extractor and retriever models to better match your specific use cases, data, and domain. By providing examples and feedback, you can improve model performance on specialized tasks without having to build models from scratch.
Domain Adaptation
Adapt general-purpose models to your specific industry, content types, or terminology
Task Optimization
Fine-tune models for specific tasks like classification, extraction, or similarity matching
Available Tuning Approaches
How Model Tuning Works
Collect Examples
Gather a set of representative examples and annotations that demonstrate the desired behavior
Create Tuning Set
Upload your examples to create a tuning dataset in Mixpeek
Configure Tuning Job
Specify which model to tune and set the tuning parameters
Launch Tuning Process
Start the tuning job and monitor progress
Evaluate Results
Compare performance metrics between the original and tuned models
Deploy Tuned Model
Apply the tuned model to your pipelines for improved performance
Best Practices
Data Quality
Focus on high-quality, diverse examples that represent the full range of cases in your domain
Sample Size
For best results, provide at least 50-100 examples per class or task, with more for complex domains
Balanced Classes
Try to provide roughly equal numbers of examples for each class or category
Iterative Approach
Tune models iteratively, evaluating performance and adding examples for challenging cases
Very small tuning sets (fewer than 20 examples) may not provide significant improvements and could lead to overfitting. For best results, provide diverse, high-quality examples.
Limitations
- Model tuning preserves the general capabilities of the base models while adapting them to your domain
- Complete retraining from scratch is not supported
- Some advanced model architectures may have limited tuning capabilities
- Tuning jobs may take anywhere from a few minutes to several hours depending on dataset size and complexity
- Maximum number of examples per tuning set is 10,000
- For very specialized domains, custom model building (available through our Professional Services) may be more appropriate
Was this page helpful?