Build an LLM Evaluation Pipeline

Your team needs to evaluate a language model before deploying it. You have a MockLLM that simulates generation and log probabilities. Your job is to build the evaluation pipeline: compute perplexity over reference texts, measure exact-match accuracy on Q&A pairs, implement LLM-as-judge scoring for open-ended responses, and wire everything into a single run_eval_suite function that returns structured results. The MockLLM and data structures are provided. You only need to implement the four metric functions.

1. 1.Compute perplexity from log probabilities: PPL = exp(-average_log_prob)
2. 2.Compute exact-match accuracy by comparing generated answers to ground truth
3. 3.Implement LLM-as-judge that constructs a rating prompt and parses a 1-5 score
4. 4.Wire all metrics into run_eval_suite and return a structured results dict
5. 5.Handle edge cases: empty inputs, parse failures in judge scores