Amazon SageMaker adds full fine-tuning to AI serverless customization

Covered by 2 sources
Read onAWS Blog· 11 min
Amazon SageMaker AI Serverless Model Customization now supports full fine-tuning, enabling improved domain-specific adaptation of large language models without requiring users to manage infrastructure. This approach updates all model parameters and has demonstrated higher accuracy in coding tasks compared to previous methods like LoRA.

What happened

Amazon Web Services has introduced full fine-tuning support for its SageMaker AI Serverless Model Customization service. This update allows users to adapt large language models by updating all model parameters, rather than just a subset. The service manages all infrastructure automatically, so users do not need to set up GPUs or handle compute resources. AWS tested the new feature on coding tasks using over 100,000 training examples.

Key details

  • Full fine-tuning updates all parameters of a model, not just a small subset as with LoRA.
  • AWS's tests achieved 90% semantic correctness in code generation tasks with full fine-tuning.
  • The Magicoder-Evol-Instruct-110K dataset was used, containing 110,000 coding instruction-response pairs.
  • SageMaker AI Serverless Model Customization automates compute allocation and checkpoint management.
  • Excessive fine-tuning can reduce a model's general knowledge.

Why it matters

Businesses and developers using SageMaker can now achieve higher accuracy on specialized tasks without managing hardware or infrastructure. This enables organizations to adapt large language models more effectively to their own data and domains, potentially improving performance in areas like code generation. However, users must balance customization with the risk of reducing the model's broader capabilities if fine-tuned too extensively.

Also covered by 1 source


More in Machine Learning