AIF-C01 notes
Developing generative AI solutions

Evaluating an FM

Evaluation checks whether the model meets the business objectives.

Three evaluation methods

MethodStrengthsWeaknesses
Human evaluationThe gold standard. Judges coherence, relevance, factuality, and qualitySlow and expensive at scale
Benchmark datasetsStandardized comparison across models and over timeMay not match your specific use case
Automated metricsQuick and scalable for fast iterationMiss nuance and may not match human judgment

Benchmarks to know:

  • GLUE: language understanding (classification, Q&A, inference)
  • SuperGLUE: harder GLUE tasks
  • SQuAD: question answering
  • WMT: machine translation

Metrics to know

MetricMeasuresBest for
ROUGEOverlap between generated text and a reference, focused on recallSummarization
BLEUn-gram precision against a referenceMachine translation
BERTScoreSemantic similarity using BERT embeddings and cosine similarityAny generation where meaning matters more than exact wording
PerplexityHow well the model predicts the next token (lower is better)Language model quality
F1 scoreBalance of precision and recallClassification and entity recognition

Automated metrics give a first read. Combine them with human evaluation for a complete picture.

On this page