Chapter 24
24Evaluating Language Models
24.1Why evaluation is the hard part
Planned: Open-ended output has no single ground truth; the metric problem.
24.2Benchmarks and their discontents
Planned: MMLU-style suites, contamination, saturation, and gaming.
24.3LLM-as-judge
Planned: Using a model to grade a model; bias, calibration, and when to trust it.
24.4Human eval and arenas
Planned: Pairwise human preference, Elo-style ranking, and their limits.
24.5Building evals for your own product
Planned: Task-specific eval sets; the flywheel of logging, labeling, and regression tests.