In testing, we tuned Haiku to act as a moderator on an example dataset. Fine-tuning improved classification accuracy from 81.5% to 99.6% and reduced prompt tokens per query by 89%. Fine-tuning for Claude 3 Haiku is available today in preview in Amazon Bedrock.
Claude 3 Haiku Fine-tuning Achieves 99.6% Classification Accuracy
By
–