Building specialized AI without sacrificing intelligence: Nova Forge data mixing in action

This article demonstrates how Amazon Nova Forge's data mixing approach enables building specialized AI models while preserving general capabilities, addressing the catastrophic forgetting problem in supervised fine-tuning.

• Nova Forge allows blending proprietary data with Amazon Nova-curated training data during fine-tuning
• Full-parameter SFT on customer VOC data improved Nova 2 Lite F1 score by 17 points (0.387 to 0.5537)
• Fine-tuning on customer data alone caused MMLU accuracy to drop from 0.75 to 0.47, indicating catastrophic forgetting
• Data mixing (75% customer + 25% Nova data) maintained MMLU accuracy at 0.74 while achieving 12-point F1 improvement
• Evaluated on complex four-level customer feedback classification with 1,420 leaf categories across 15,372 samples
• Qwen3-30B model showed severe degradation after fine-tuning, losing instruction-following ability entirely

Nova Forge's data mixing effectively balances domain specialization with general capability retention, enabling robust enterprise AI deployment without sacrificing foundational model intelligence.