AI Model Customization Job Submission

AI Model Customization Job Submission#

SageMaker Python SDK V3 provides specialized trainer classes for different model customization approaches, along with advanced features like data mixing, recipe overrides, evaluation, and deployment.

Compute Options#

By default, training jobs run on serverless compute — fully managed infrastructure that abstracts away instance provisioning and scaling. You do not need to specify a compute parameter to use serverless.

If you need dedicated instances, you can optionally specify:

Trainer Classes#

SFTTrainer (Supervised Fine-Tuning)

Traditional fine-tuning with labeled datasets for task-specific adaptation

CPTTrainer (Continued Pre-Training)

Continue pre-training on a raw corpus to extend model knowledge in a specific domain

DPOTrainer (Direct Preference Optimization)

Fine-tune models using human preference data without reinforcement learning complexity

RLAIFTrainer (Reinforcement Learning from AI Feedback)

Use AI-generated feedback to improve model behavior and alignment

RLVRTrainer (Reinforcement Learning from Verifiable Rewards)

Fine-tune with verifiable reward signals for objective optimization

MultiTurnRLTrainer (Agentic Reinforcement Fine-Tuning)

Fine-tune models for multi-turn agent interactions using reinforcement learning from environment feedback

Key Features#

Data Mixing (Nova, serverless only)

Blend your custom data with Nova’s curated synthetic datasets (code, math, chat, planning, reasoning, etc.) to prevent catastrophic forgetting while specializing on your domain. See Nova Data Mixing for details.

Recipe Overrides

Layer training configuration from multiple sources (YAML recipes, override dicts, SDK defaults) with clear precedence. Use get_resolved_recipe() to inspect the merged configuration before job submission. See Fine-Tuning with Serverful Training Jobs and Fine-Tuning with HyperPod for examples.

Customization Techniques

Contents