Chapter 10: Evaluation & Benchmarks

Evaluating language models is fundamentally harder than evaluating classifiers or regressors. There is no single "accuracy" metric for open-ended text generation. A model's response to "Explain quantum computing" could be correct in countless different ways, and automated metrics can only capture limited aspects of quality.

The evaluation landscape spans a spectrum from fully automated to fully human. Automated metrics like perplexity measure how well a model predicts text. Task-specific metrics like BLEU and ROUGE compare generated text to reference outputs. LLM-as-judge uses a stronger model to evaluate a weaker one. Human evaluation captures nuances that no automated metric can.

Benchmarks provide standardized tests across capabilities: reasoning (GSM8K), knowledge (MMLU), coding (HumanEval), safety (TruthfulQA), and more. But benchmarks have limitations -- models can be trained on benchmark data, and high benchmark scores don't always translate to real-world usefulness.

Red teaming is adversarial evaluation: deliberately trying to make the model fail, produce harmful content, leak information, or behave unexpectedly. This is essential before deployment because users will inevitably push models beyond their intended use cases.

This chapter covers:

• Perplexity: The fundamental metric for language model quality
• BLEU & ROUGE: Automated metrics for translation, summarization, and text generation
• Human Evaluation: When and how to use human judges effectively
• Benchmark Suites: Standardized tests across reasoning, knowledge, coding, and safety
• Red Teaming: Adversarial testing to find failure modes and safety vulnerabilities