Fine-tuning
Train your local LLM on your DeFi voice commands. Fully private, fully on-device.
Training Data
Voice commands collected from your DeFi interactions
Checking training data...
LoRA Configuration
Hyperparameters for the fine-tuning run
LoRA (Low-Rank Adaptation) injects small trainable weight matrices into the LLM's attention layers. Training is fast and memory-efficient. Higher rank = more capacity, more VRAM. Alpha controls the scaling factor applied to the LoRA updates.
Fine-tuning Controls
Start, pause, or cancel the training run
What is Fine-tuning?
How Cortex learns from your on-chain behavior
QVAC LoRA Fine-tuning
Automatic data collection — Cortex collects training data from every successful voice command — intents, parsed parameters, and execution results. You never have to label anything manually.
LoRA adapts to your vocabulary — Low-Rank Adaptation fine-tunes the LLM to understand your specific DeFi vocabulary, wallet addresses, and interaction patterns without retraining the full model.
Fully local via QVAC — All training runs on your device through the QVAC SDK. No data leaves your machine. No cloud compute. No API keys.
Auto-loaded on restart — Once training completes, the LoRA adapter is saved locally and loaded automatically the next time Cortex starts. Your model gets smarter with every session.
LoRA rank 8 with alpha 16 is a solid starting point. Increase rank for more expressive capacity if you have many training examples and sufficient VRAM.