Evaluating an LLM means measuring the quality of its answers on representative test sets: accuracy on tasks with known answers, instruction following, faithfulness to sources, safety. You combine automatic metrics, evaluation by a judge model and human review, keeping in mind their respective biases.