Fine-tuning models

Customizing your models

How FMs benefit from fine-tuning

Customize vs. augment

Fine-tuning vs. Retrieval Augmented Generation (RAG)

Fine-tuning

• Is specific style, behavior, or vocabulary required?
• Is training data available?
• Do you need to reduce the risk of hallucinations?

RAG

• Is knowledge from external data sources required?
• Is data dynamic or changing?
• Do you need to know the sources of answers?

Custom models in Amazon Bedrock

Components of a customization job

Customization architecture overview

Security and privacy

• You are always in control of your data
• Data not used to improve models, and not shared with model providers
• Customer data remain in Region
• Support for AWS PrivateLink and VPC configurations
• Integration with AWS IAM
• API monitoring in AWS CloudTrail, logging & metrics in Amazon CloudWatch
• Custom models encrypted and stored with Service or Customer Managed Keys (CMK) - Only you have access to your models

Customizing model responses for your business

Datasets for instruction

Model fine-tuning

• Place in S3 your dataset for train, and validation
• jsonl format comprised of instructions { ”prompt": ”input text”, ”completion": ”output text” } Up to 10k training records, and 1k validation records
• Set hyperparameters
• Epoch: 1-10
• Batch size: defaults to 1
• Learning rate: defaults to 0.00005
• Learning warmup steps: recommended 0

Model continued pre-training

• Place in S3 your dataset for train, and validation
• jsonl format comprised of unlabeled data { ”input": ”input text”} Up to 100k training records, and 1k validation records
• Set hyperparameters
• Epoch: 1-10
• Batch size: defaults to 1
• Learning rate: defaults to 0.00005
• Learning warmup steps: recommended 0