rapidrouge

⚠️⚠️⚠️ This repository was build with AI ⚠️⚠️⚠️

A fast, pure-Python drop-in replacement for rouge-score.

uv add rapidrouge          # pure-Python, ZERO runtime deps — installs anywhere

Optional extras: rapidrouge[stemmer] (nltk, for use_stemmer=True), rapidrouge[aggregate] (numpy, for BootstrapAggregator), or rapidrouge[full].

Same import, same results as rouge_score 0.1.2 — ROUGE-L's LCS length is computed with the Hyyrö bit-parallel algorithm ([1], [2], [3]), so it's far faster on long documents:

from rapidrouge import rouge_scorer
scorer = rouge_scorer.RougeScorer(["rouge1", "rougeL"], use_stemmer=False)
scorer.score("the quick brown fox", "quick brown the fox")["rougeL"].fmeasure  # 0.75

Benchmarks

ROUGE-L is where the bit-parallel kernel pays off, and the win grows with document length. From ~10× on a sentence to ~900× on a 35k-token document (115 ms vs ~104 s), with results identical to rouge_score:

It's a memory win too: the reference fills an O(n·m) DP table (~9.8 GB at 35k tokens), while rapidrouge's LCS is O(n). rouge1/ rouge2 (n-gram counting) and rougeLsum (still a DP table) match the reference in both correctness and speed (~1×), as expected — only ROUGE-L rides the Hyyrö kernel.

License

Apache-2.0. Portions derived from rouge_score (Apache-2.0); see NOTICE.