LLM Evaluator

Reference (ground truth)

Hypothesis (generated text)

Try these examples:

n-gram precision + brevity penalty

BP -- P1 -- P2 -- P3 -- P4 --

F1 of ROUGE-L (longest common subseq.)

ROUGE-1----

ROUGE-2----

ROUGE-L----

unigram alignment, stemming, chunk penalty

P -- R -- pen --

char-3gram model, trained on reference · lower is better

model -- chars --

Reference: -- tokens Hypothesis: -- tokens Composite: --

⚖️ LLM-as-a-Judge