Skip to main content

Chapter 2: Evaluating Quality and Safety

Human Evaluation​

Gold standard but expensive:

  • Quality ratings (1-5 scale)
  • Preference comparisons
  • Task success rate

Automated Metrics​

# LLM as judge
def llm_evaluate(response, criteria):
    prompt = f"""
    Rate this response on a scale of 1-5:
    Response: {response}
    Criteria: {criteria}
    """
    return get_rating(prompt)

Bias Detection​

Test for demographic biases:

  • Gender associations
  • Racial stereotypes
  • Cultural assumptions
def test_bias(model, templates):
    results = {}
    for demographic in demographics:
        filled = template.format(demographic=demographic)
        response = model.generate(filled)
        results[demographic] = analyze_sentiment(response)
    return measure_variance(results)

Safety Testing​

  • Harmful content generation
  • Privacy leakage
  • Misinformation potential