# Property Finder fixes evaluation bias to cut Amazon Bedrock vision costs 86 percent

> Original title: Fix evaluation bias to unlock better models on Amazon Bedrock

- Source: AWS Machine Learning Blog
- Published: 2026-10-27T16:53:02.000Z
- Bad Influence: https://badinfluence.news/items/ljzir3ichl2g2tmuyob20u2cp
- Original: https://aws.amazon.com/blogs/machine-learning/fix-evaluation-bias-to-unlock-better-models-on-amazon-bedrock/

## Summary

Property Finder found that 96.8 percent of its evaluation labels came from its incumbent vision model, which made alternatives look worse. After building expert-annotated ground truth and tuning prompts, it switched to Llama 4 Maverick 17B Instruct on Amazon Bedrock. The post reports Micro F1 rising from 0.55 to 0.81 on the development set, with estimated costs down 86 percent.

## Why it matters

The post shows how biased evaluation labels can hide a better model, and how to rebuild ground truth and tune prompts for a vision workload.