What Happened

Amazon Web Services published the second installment of its Nova Forge SDK tutorial series on the AWS Machine Learning Blog, providing a hands-on walkthrough for fine-tuning Amazon Nova models using data mixing techniques. The guide covers a five-stage workflow from environment setup through model evaluation, targeting engineers with access to SageMaker HyperPod infrastructure.

The post builds directly on part one of the series, which introduced the SDK and covered initiating customization experiments. Part two focuses exclusively on data mixing — the technique of blending customer-specific training data with Amazon-curated datasets during supervised fine-tuning.

Why It Matters

The practical case for data mixing is quantified in the post. According to AWS, blending customer data with Amazon-curated datasets during fine-tuning delivered a 12-point F1 improvement on a Voice of Customer classification task spanning 1,420 leaf categories, while preserving near-baseline Massive Multitask Language Understanding (MMLU) scores. AWS states that fine-tuning an open-source model on customer data alone caused a near-total loss of general capabilities — a well -known catastrophic forgetting problem that data mixing is designed to address.

For engineering teams evaluating whether to fine-tune proprietary models versus open-source alternatives on AWS , this benchmark comparison — though limited to a single classification task — provides a concrete data point on the tradeoff. The 1,420- category classification scope also signals the technique's applicability to large -scale enterprise taxonomy problems.

The Technical Detail

The workflow AWS documents runs across five stages:

  • Environment setup: Install the Nova Forge SDK and configure AWS resources. The SDK requires SageMaker Hy perPod CLI tooling, distributed via a private S3 bucket provided during Nova Forge onboarding.
  • Data preparation: Load, sanitize, transform, validate, and split training data.
  • Training configuration: Configure the SageMaker HyperPod runtime, MLflow experiment tracking, and data mixing ratios.
  • Model training : Launch a supervised fine-tuning job using Low-Rank Adaptation (LoRA) and monitor via HyperPod.
  • Model evaluation: Run public benchmarks and domain-specific evaluations against the fine-tuned checkpoint.

The reference hardware configuration uses ml.p5.48 xlarge instances — AWS's highest-tier GPU compute nodes — with four instances allocated for both training and evaluation. AWS explicitly recommends validating configuration with a short test run using max_steps=5 before committing to a full training job, acknowledging the cost exposure of this instance class.

Fine-tuning uses LoRA (Low-Rank Adaptation), which reduces trainable parameter count relative to full fine-tuning — a relevant constraint when working on HyperPod at ml.p5.48xlarge scale. Experiment tracking is handled through Amazon SageMaker MLflow, keeping the tool chain within the AWS ecosystem.

Access to Nova Forge remains g ated: the SDK distribution bucket and onboarding credentials are provided separately, meaning teams cannot self-serve without prior AWS enrollment in the program. Required IAM permissions span SageMaker, Amazon S3, and Amazon CloudWatch.

Infrastructure Prerequisites

Standing up the environment requires:

  • An AWS account with Nova Forge access
  • A SageMaker HyperPod cluster on Amazon EKS with provis ioned GPU compute nodes and execution roles
  • A SageMaker MLflow application
  • An IAM role with SageMaker, S3, and CloudWatch permissions
  • SageMaker Studio notebook or equivalent Jupyter environment

The EKS dependency adds non -trivial setup overhead for teams not already running H yperPod workloads. AWS links to the HyperPod getting-started documentation for cluster provisioning details, but does not inline those steps in this tutorial.

What To Watch

  • Part three of the Nova Forge SDK series: AWS has structured this as an ongoing series. The next installment will likely address evaluation methodology, deployment, or advanced mixing strategies — watch the AWS ML Blog for the follow-up, expected within weeks based on the part one to part two cadence.
  • Nova Forge general availability: The SDK currently requires onboarding through AWS, suggesting it remains in a controlled access phase. Any move toward broader or self -serve availability would materially lower the barrier for enterprise adoption.
  • Competitive fine-tuning tooling: Google's Vertex AI and Azure AI Studio both offer managed fine-tuning pipelines. AWS publishing detailed benchmarks on catastrophic forgetting mit igation adds pressure on competitors to publish equivalent data mixing results for their own hosted model fine-tuning offerings.
  • LoRA configuration specifics: The post references LoRA as the fine-tuning method but does not publish rank, alpha, or dropout parameters used in the 12-point F1 result. Subsequent posts or community reproduction attempts may surface those details.
  • Pricing transparency: AWS defers to the SageMaker pricing page for ml.p5.48xlarge costs rather than quoting figures in the post. At current on-demand rates for p 5 instances, four-instance training runs represent meaningful spend — watch for any Nova Forge-specific pricing announcements.