Stop 18 · Evaluation, Safety & Responsible AITopics · 1/9

LLM Evaluation Metrics (BLEU, ROUGE, Perplexity)

In short

You changed the prompt. Is the output better or just different? Automatic metrics give you a quick number, as long as you know what each one can and can't see.

Automated metrics for scoring generated text against references — useful as a fast signal but each has real blind spots that make purely automated evaluation insufficient alone.

The blob as an examiner ticks a scribbled answer sheet with a red pen, a second sheet beside it.

Think of it like this

The keyword-counting examiner

Some examiners check answers by ticking keywords from the model answer. Write 'mitochondria' and 'energy' and you get marks, even if your sentence is nonsense. BLEU and ROUGE grade like this examiner: they count matching words, not meaning.

Watch it happen

Counting word overlap with a reference

The

Pune

train

departs

1/4Reference: 'The train to Pune leaves at 6 pm'. We check which words of the model's output also appear in it.

The big ideas

What you need to know

BLEU: precision of n-grams

How many of the output's word chunks appear in the reference. Built for machine translation.

ROUGE: recall of n-grams

How much of the reference shows up in the output. ROUGE-L uses the longest common word sequence. Built for summaries.

Perplexity: how surprised

How well a model predicts real text, from its own token probabilities. Lower is better. It says nothing about helpfulness.

Numbers aren't meaning

A correct answer in new words scores low. A wrong answer with the right words scores high. Use these as quick signals only.

See it in code

Walk through the code

python

1import sacrebleu2from rouge_score import rouge_scorer3 4ref = "The train to Pune leaves at 6 pm from platform 3."5hyp = "The Pune train departs at 6 pm from platform 3."6 7bleu = sacrebleu.sentence_bleu(hyp, [ref])8print("BLEU:", round(bleu.score, 1))9 10scorer = rouge_scorer.RougeScorer(["rouge1", "rougeL"], use_stemmer=True)11scores = scorer.score(ref, hyp)12print("ROUGE-1 F1:", round(scores["rouge1"].fmeasure, 3))13print("ROUGE-L F1:", round(scores["rougeL"].fmeasure, 3))

1/4A reference answer and the model's answer. Same meaning, different words.

BLEU with sacrebleu and ROUGE with rouge-score. Both need a human-written reference answer to compare against.

Save yourself the pain

Common mistakes

Oops

Using BLEU to judge a chatbot's open-ended answers.

Do this instead

Many good answers exist for an open question. Use rubric scoring, LLM judges or human review instead.

Oops

Comparing perplexity across models with different tokenizers.

Do this instead

Perplexity depends on the tokenizer. Only compare models that split text the same way, on the same data.

Your mission

Break the metrics

~20 min

Quick check

A summary is factually wrong but copies many phrases from the reference. What will ROUGE likely show?

Go deeper

1Compute BLEU/ROUGE scores for a small set of generated outputs against references.
2Learn what each metric actually measures (n-gram overlap) and what it misses (meaning, fluency).
3Understand why perplexity measures a model's confidence, not correctness.