LLM Evaluation Cheat Sheet

Comprehensive guide to evaluating LLM outputs — perplexity, BLEU, ROUGE, human evaluation, lm-eval-harness, and benchmark frameworks.

Last Updated: May 1, 2025

Core Metrics

Metric	Measures	Range	Best For
Perplexity (PPL)	How 'surprised' model is by text	1 to ∞ (lower=better)	Language modeling quality
BLEU	N-gram overlap with reference	0–100 (higher=better)	Translation, summarization
ROUGE-1/2/L	Recall-oriented n-gram overlap	0–100 (higher=better)	Summarization recall
METEOR	BLEU + synonyms + stemming	0–100 (higher=better)	Translation with paraphrasing
BERTScore	Semantic similarity via embeddings	0–1 (higher=better)	Captures meaning, not exact words
Exact Match (EM)	Exact string match	0 or 1	QA with short answers
F1 Score	Token-level precision + recall	0–100 (higher=better)	QA, NER, classification
chrF	Character n-gram F-score	0–100 (higher=better)	Morphologically rich languages

Popular Benchmarks

Benchmark	What It Tests	Dataset Size	Notable Models
MMLU	57 subjects — knowledge + reasoning	~14K questions	GPT-4 86.4%, Claude 3 88.7%
HellaSwag	Commonsense reasoning	~10K completions	Tests grounded reasoning
HumanEval	Python code generation	164 problems	Code-focused evaluation
MT-Bench	Multi-turn conversation quality	80 questions × 2 turns	GPT-4 judges responses
AlpacaEval	Instruction following	805 prompts	Win rate vs reference (GPT-4)
BigBench	200+ diverse reasoning tasks	Varies by task	Broad cognitive evaluation
TruthfulQA	Factual accuracy vs misconceptions	817 questions	Measures hallucination tendency
GSM8K	Grade-school math word problems	8.5K problems	Multi-step math reasoning

Running lm-eval-harness

pip install lm-eval

Install the EleutherAI evaluation harness

lm_eval --model hf --model_args pretrained=meta-llama/Llama-2-7b-hf --tasks mmlu --num_fewshot 5

Run MMLU with 5-shot prompting on Llama-2

lm_eval --model hf --model_args pretrained=gpt2 --tasks hellaswag,arc_easy --batch_size 8

Evaluate on multiple tasks in one run

lm_eval --model openai --model_args model=gpt-4 --tasks gsm8k --num_fewshot 8

Evaluate OpenAI models via API

lm_eval --model vllm --model_args pretrained=model-name,tensor_parallel_size=2 --tasks mmlu

Use vLLM backend for faster inference

--output_path ./results

Save evaluation results to a directory

--limit 100

Limit evaluation to 100 examples per task (quick test)

--log_samples

Log individual predictions alongside aggregate metrics

Human Evaluation Methods

Item	Description
`Pairwise Comparison`	Humans compare two model outputs side-by-side — used in LMSys Chatbot Arena
`Likert Scale Rating`	Rate outputs 1-5 on quality, relevance, safety, and helpfulness
`A/B Testing`	Deploy two models to production and measure user engagement/satisfaction
`Vibe Check`	Qualitative assessment by domain experts — essential for creative/ nuanced tasks
`Red Teaming`	Adversarial testing — humans deliberately try to elicit unsafe or incorrect outputs
`Inter-Annotator Agreement`	Measure consistency between human judges (Cohen's kappa, Fleiss' kappa)
`Elo Rating System`	Derive rankings from pairwise battles (used by Chatbot Arena)
`LLM-as-Judge`	Use GPT-4 to evaluate other models — correlates well with human judgment

Pro Tip: Never rely on a single metric. Combine automatic metrics (perplexity/BLEU/ROUGE) with human evaluation and task-specific benchmarks for a complete picture of model quality.