For LLMs, standard metrics don't work: you can't just check exact match. BLEU, ROUGE, perplexity, and BERTScore each capture different aspects of quality. Edit outputs below and see the scores update live.
Sign in to save progress
Edit the model output and watch BLEU, ROUGE-1, ROUGE-2, and word overlap update in real time.
BLEU
Clipped n-gram precision + brevity penalty (computed here up to bigrams)
ROUGE-1
Unigram recall overlap
ROUGE-2
Bigram recall overlap
BERTScore
Semantic similarity via embeddings
Tip: try making the output shorter, longer, or rephrasing it to see how each metric reacts.
Three classic failure modes where automated metrics mislead you. Live scores are computed from the actual text.
Reference
The dog bit the man.
Model output
The man bit the dog.
Same words, opposite meaning. Every word matches and three of the four bigrams survive the swap, so BLEU-2 stays high while the semantics are reversed. Only longer n-grams would expose it.
Reference
The capital of France is Paris.
Model output
Paris is the French capital.
A perfect paraphrase is penalised because the word order differs: not a single bigram matches, so BLEU-2 collapses to zero. BLEU punishes creativity.
Reference
The model generates coherent and accurate text.
Model output
The the the the the the the.
Unclipped precision would score this 1.00, since every "the" appears in the reference. Clipping caps it at the reference count, so BLEU collapses to zero. Perplexity can also be fooled: a model stuck in a repetition loop grows ever more confident in the repeated token, so low perplexity alone never guarantees quality.
The four benchmarks every LLM paper cites, and what they actually measure.
MMLU
Massive Multitask Language Understanding
General knowledge across 57 academic subjects
Multiple-choice questions spanning STEM, humanities, law, medicine
Reference score (at publication)
GPT-4o (2024): 88.7%
HumanEval
HumanEval
Code generation: 164 Python problems
Pass@1 rate: model writes a function, tests run against hidden test cases
Reference score (at publication)
GPT-4o (2024): 90.2%
HellaSwag
HellaSwag
Commonsense reasoning & sentence completion
Model picks the most plausible next sentence from 4 adversarially filtered choices
Reference score (at publication)
GPT-4 (2023): 95.3%
MT-Bench
Multi-Turn Benchmark
Multi-turn conversation quality
GPT-4 judges responses on writing, reasoning, math, coding across 80 multi-turn questions
Reference score (at publication)
GPT-4 (2023): 8.99 / 10
Each approach has trade-offs. The emerging winner is LLM-as-judge: using a stronger model to score a weaker one.
Human Eval
Automated Metrics
LLM-as-Judge
Key Insight
LMSYS Chatbot Arena ranks models with Elo-style ratings computed from millions of blind, head-to-head human votes. Static academic benchmarks can leak into training data and be gamed; live human preference is much harder to optimize against, which is why many teams treat Arena rankings as the stronger signal.