Adapters
LoRA fine-tuning with training profiles and managed lifecycle.
Adapters allow you to fine-tune language models using LoRA (Low-Rank Adaptation) to create domain-specific versions that perform better on your particular use cases. MechaMental manages the entire lifecycle from training data to deployment.
Training Profiles
MechaMental offers three training profiles that balance cost, speed, and quality:
| Profile | Speed | Cost | Best For |
|---|---|---|---|
| Economy | Slower | Low | Experimentation and initial prototyping |
| Standard | Moderate | Medium | Production-quality adapters with reasonable turnaround |
| Express | Fast | High | Time-sensitive updates and rapid iteration |
Choose a profile based on your needs. Economy is ideal for testing whether fine-tuning helps your use case before committing to a full training run. Express is best when you need to iterate quickly on adapter quality.
Adapter Lifecycle
Create
Define the adapter with a base model, training profile, and training data. The training data format depends on the base model but typically consists of input/output pairs that demonstrate the desired behavior.
Train
Submit the training job. Training runs asynchronously on managed infrastructure. You can monitor progress from the adapter detail page.
Evaluate
Review training metrics and test the adapter's output quality. Compare results against the base model to measure improvement on your target tasks.
Deploy
Make the adapter available as a model target in your pipelines. Once deployed, the adapter appears alongside other models in the inference step configuration.
Continuous Learning
Adapters can be retrained with new data as it becomes available. This enables continuous learning loops where your AI models improve over time based on real-world interactions and feedback.
Incremental Improvement
Each training iteration builds on the previous adapter. As you collect more domain-specific examples from production usage, you can retrain to steadily improve output quality without starting from scratch.
Using Adapters in Pipelines
Once deployed, an adapter appears as a model target that can be referenced in inference steps. You can use it as the primary model or as a fallback in a model's fallback chain. This lets you try the fine-tuned adapter first and fall back to the base model if needed.