AIF-C01 notes
Optimizing foundation models

Evaluate Results

The two most common ways to evaluate a generative AI model, and why you usually use both:

Human evaluationBenchmark datasets
KindQualitativeQuantitative and objective
MeasuresUser experience, contextual appropriateness, creativity and flexibilityAccuracy, speed and efficiency, scalability
Best forIterative tuning to meet user expectationsInitial testing and comparing models or model versions

Building a benchmark dataset

  1. Subject matter experts (SMEs) write relevant, challenging questions about the topic or documents.
  2. SMEs provide the correct answers and the context they come from.
  3. The model answers the questions and its answers are scored against the SMEs' answers.

Scoring can be automated with LLM as a judge: another LLM compares the model's answers to the benchmark answers.

Combined approach

Benchmarks show technical capability, and humans show real-world usefulness. AnyCompany tests against benchmarks before production and collects human ratings after, so the model keeps improving.

On this page