Building specialized AI without sacrificing intelligence: Nova Forge data mixing in action
New AWS service blends proprietary data with curated training to preserve general intelligence.
Amazon has launched Nova Forge, a new AWS service designed to solve a critical enterprise AI problem: catastrophic forgetting. When companies fine-tune large language models (LLMs) like GPT-4 or Claude on their proprietary data, the models often lose their general reasoning and instruction-following abilities. Nova Forge addresses this by allowing developers to start from early model checkpoints and blend their proprietary data with Amazon's curated Nova training data, enabling specialization without sacrificing the broad intelligence that makes foundation models useful across an organization.
In a comprehensive evaluation by the AWS China Applied Science team, Nova Forge was tested on a challenging real-world Voice of Customer (VOC) classification task with over 16,000 samples across a 1,420-category hierarchy. The service delivered a 17% improvement in the domain-specific F1 score while successfully preserving the model's general capabilities, as measured by near-baseline Massive Multitask Language Understanding (MMLU) scores. This data mixing methodology provides a practical path for enterprises to deploy a single, versatile model that can handle specialized workflows like customer feedback routing while still performing general tasks like analysis and documentation.
- Solves catastrophic forgetting: Maintains near-baseline MMLU scores after fine-tuning, preserving general reasoning.
- Delivers 17% F1 score improvement on a complex, 1,420-category customer feedback classification task.
- Uses a novel 'data mixing' approach, blending proprietary enterprise data with Amazon's curated Nova training datasets.
Why It Matters
Enables enterprises to deploy a single, versatile AI model that is both a domain expert and a generalist, eliminating costly trade-offs.