Nova Forge SDK series part 2: Practical guide to fine-tune Nova models using data mixing capabilities

This article provides a comprehensive guide to fine-tuning Amazon Nova models using the Nova Forge SDK with data mixing capabilities, enabling domain specialization while preserving general intelligence.

• Data mixing blends domain-specific training data with Amazon-curated datasets to prevent catastrophic forgetting
• Five-stage workflow: environment setup, data preparation, training configuration, model training, and evaluation
• Requires SageMaker HyperPod cluster with GPU instances (ml.p5.48xlarge recommended)
• Data sanitization removes tokens conflicting with Nova's chat template (e.g., "System:", "Assistant:")
• LoRA (Low-Rank Adaptation) is parameter-efficient; full-rank SFT available for deeper adaptation
• Configurable data mixing ratios control batch composition (e.g., 50% customer data, 50% Nova-curated)
• Evaluation measures both domain performance and general capability retention using MMLU and IFEval benchmarks
• Previous results showed 12-point F1 improvement on domain task while maintaining near-baseline MMLU scores
• Best practice: iterate on mixing ratios, not just hyperparameters; always evaluate on both axes

The guide demonstrates that data mixing enables practical production fine-tuning by balancing domain expertise with general intelligence through systematic iteration and dual-axis evaluation.