Training Engine

The Training Engine is what makes QuantumVerifi improve over time. After each analysis, Scout harvests what worked and what didn’t, then uses that data to fine-tune a model specific to your organisation.

Available on Scale and Enterprise plans.

How It Works

1. Harvest — After each analysis, successful test patterns, self-healing corrections, and validation outcomes are collected into a training dataset
2. Train — When enough data accumulates, a QLoRA adapter is fine-tuned on a base code model (e.g., Qwen2.5-Coder-7B)
3. Activate — The trained adapter is activated for your tenant, routing future generation requests through your custom model
4. Infer — New analyses use your adapter for context-aware test generation

What Improves

With a trained adapter, Scout generates tests that:

• Match your codebase’s naming conventions and test patterns
• Use the correct imports and framework idioms from the start
• Require fewer self-healing cycles (tests pass on first attempt more often)
• Cover edge cases specific to your application domain

Training Dashboard

The training dashboard at Settings > Training shows:

• Active adapters — which operations are using your custom model
• Training jobs — status, metrics, and hyperparameters for each training run
• Metrics — routing statistics showing how often your adapter is used

Adapter Operations

Each adapter can be activated for specific operations:

You can activate an adapter for one or more operations, and deactivate it at any time.

Training Requirements

• Minimum data: At least 5 completed analyses for meaningful training
• Compute: Training runs on GPU infrastructure (managed in cloud, or your own GPUs for self-hosted)
• Duration: A typical training run takes 10-30 minutes depending on dataset size
• Storage: Trained adapters are stored as compact LoRA weights (typically 50-200 MB)

What’s Next

• How Training Works — Technical details of the 4-phase training flywheel