rougers

ROUGE metrics in Rust, with scores identical to Google's rouge-score, about 30 times faster on one core and over 100 times faster on all of them.

rouge-score is the reference implementation behind Hugging Face evaluate, most summarization papers and many LLM evaluation pipelines. It is written in pure Python and becomes the bottleneck once you score tens of thousands of outputs. rougers reproduces it exactly, down to the last bit of every float, including the tokenizer, NLTK's Porter stemmer and the summary-level rougeLsum.

The Rust crate is rougers; the Python package built on it is rouge-score-rs. This is an independent project, not affiliated with Google or the authors of rouge-score.

Python

pip install rouge-score-rs

Change the import and keep the rest of your code:

from rouge_score_rs import rouge_scorer  # was: from rouge_score import rouge_scorer

scorer = rouge_scorer.RougeScorer(["rouge1", "rouge2", "rougeL", "rougeLsum"], use_stemmer=True)
scores = scorer.score(
    "The quick brown fox jumps over the lazy dog",
    "The quick brown dog jumps on the log.",
)
scores["rouge1"]
# Score(precision=0.75, recall=0.6666666666666666, fmeasure=0.7058823529411765)

score_multi, custom tokenizers, split_summaries, scoring.BootstrapAggregator and tokenize.tokenize work the same way as in rouge-score, and scorers can be pickled for use in worker processes.

To score many pairs, use score_batch. It runs on all CPU cores and releases the GIL:

results = scorer.score_batch(references, predictions)  # list of dicts, one per pair

Set RAYON_NUM_THREADS to limit the number of threads.

Rust

[dependencies]
rougers = "0.1"

use rougers::{RougeScorer, RougeType};

let scorer = RougeScorer::new(["rouge1", "rougeL"], true)?;
let scores = scorer.score(
    "The quick brown fox jumps over the lazy dog",
    "The quick brown dog jumps on the log.",
);
let rouge_l = scores.get(RougeType::L).unwrap();
println!("{:.4}", rouge_l.fmeasure);

RougeScorer is Send + Sync, so it can be shared across threads, for example with Rayon.

Performance

10,000 pairs of English text from the NLTK Gutenberg corpus, about 60 words each, with rouge1, rouge2, rougeL, rougeLsum and stemming. Measured on an Apple M1 Pro with Python 3.12:

The script asserts that all three produce the same scores. To reproduce:

cd bindings/python
python benchmarks/compare.py --pairs 10000

Compatibility

The test suite compares rouge-score-rs with rouge-score and fails on any difference in any float:

• the stemmer against NLTK on all 235,976 words of a system dictionary (NLTK 3.8–3.10 stem them identically);
• tokenization against Python's str.lower for every Unicode code point;
• thousands of random texts with repeated words, punctuation, digits, newlines and non-Latin scripts, plus property-based tests on arbitrary Unicode;
• score_multi, custom tokenizers, split_summaries and bootstrap aggregation with a fixed NumPy seed.

Known differences:

• An invalid ROUGE type raises ValueError when the scorer is created, not on the first call to score. Type names must be exactly rouge1…rouge9, rougeL or rougeLsum.
• Texts must be str; rouge-score also accepts UTF-8 bytes.
• BootstrapAggregator needs NumPy (pip install 'rouge-score-rs[aggregate]') and split_summaries=True needs NLTK (pip install 'rouge-score-rs[sentences]'). Without them, rouge-score-rs has no Python dependencies.

Development

cargo test -p rougers
cd bindings/python
uv sync --group dev
uv run pytest

License

Apache-2.0, the same licence as rouge-score and NLTK, whose scoring, tokenization and stemming logic this project ports. See NOTICE.