fitzyracing-rouge (rouge)

Part of 360 Bench, tested fixes for abandoned PyPI packages.

This is a fork of rouge by pltrdy and its contributors, published as a drop-in replacement. The upstream package (~640k downloads a month, widely used to score summaries and LLM output) has not been released since 1.0.1 (July 2021). The import name is still rouge, so no code changes are needed. All credit for the original library goes to its author and contributors; it remains available under the same Apache-2.0 license (see LICENSE and NOTICE).

If a fixed rouge release appears on PyPI, prefer it and switch back.

What's fixed in this fork

Based on upstream rouge==1.0.1 (tag 1.0.1). The API is unchanged. Scores are identical to 1.0.1 except for the inputs described in the first item.

• Unrelated texts no longer share a phantom empty word. rouge splits each text on periods and dropped empty segments before collapsing whitespace, so a segment made only of whitespace (the " " in "cat. ", or between the periods in "the cat. . dog") survived as an empty sentence and was counted as an empty "" word. Two texts that both had one then "overlapped" on it:

  from rouge import Rouge
  Rouge().get_scores("the cat. . dog", "a bird. . fish")[0]["rouge-1"]
  # 1.0.1: {'r': 0.25, 'p': 0.25, 'f': 0.249...}
  # fork:  {'r': 0.0, 'p': 0.0, 'f': 0.0}
  Rouge().get_scores("cat. ", "dog. ")[0]["rouge-1"]
  # 1.0.1: {'r': 0.5, 'p': 0.5, 'f': 0.499...}
  # fork:  {'r': 0.0, 'p': 0.0, 'f': 0.0}
  

  The same happened to an empty model output (" ") scored against a reference that ends with ". ". Text with a space or newline after its last period is common in LLM output, so this quietly inflated ROUGE in evaluations. The fork drops whitespace-only segments.

  Exactly which scores change: only pairs where the hypothesis or the reference has a whitespace-only segment next to real words, that is whitespace (spaces, tabs, newlines, Unicode spaces) directly after a period and followed by another period or the end of the text. Those pairs now get exactly the scores they would get without the blank segment ("the cat. . dog" scores like "the cat. dog", "cat. " like "cat"). All other inputs score exactly as in 1.0.1 (checked against 1.0.1 on 30,000 random text pairs). Texts made only of whitespace and periods (for example an empty LLM generation " ") are scored exactly as in 1.0.1 too, rather than raising ValueError: Hypothesis is empty. like "" does, so evaluation loops keep running. return_lengths values don't change.

  Upstream issue: #77; unmerged PR: #78. The fork applies #78's core change (filter segments by strip()), but leaves out its changes to ignore_empty (skipping whitespace-only pairs, returning [] or raising ValueError when nothing is left), and keeps whitespace-only texts working.

• ROUGE-L no longer crashes with RecursionError on long sentences. The longest common subsequence was rebuilt recursively, one call per word, so a single sentence of roughly 700 to 1,000 words (for example an LLM answer without periods) exceeded Python's recursion limit. This is upstream PR #69 by @Lppy, which the maintainer merged in November 2024 but never released; it is included unchanged (commit 4b57032). It gives identical results (checked on 40,000 random cases).

• Packaging: pyproject.toml replaces setup.py. The same modules are installed as before, including the rouge command. rouge.__version__ is "1.0.2".

The three upstream tests in tests/test_basic.py already fail on 1.0.1 (their expected scores date from 2017 and predate later upstream changes); they are unchanged.

Install

pip install fitzyracing-rouge

Switching from rouge

This distribution installs the same rouge package and rouge command as the original, so your code keeps doing from rouge import Rouge. Uninstall the original first, then install the fork:

pip uninstall -y rouge
pip install fitzyracing-rouge

The order matters. Both distributions own the same files, so if you install the fork first and uninstall rouge afterwards, pip deletes the shared files and the import breaks. If that happens, run pip install --force-reinstall --no-deps fitzyracing-rouge.

In requirements.txt / pyproject.toml, replace rouge with fitzyracing-rouge. (This is pltrdy/rouge. Google's rouge-score package, imported as rouge_score, is a different project and is not affected.)

If you get rouge through another package

pip cannot replace a dependency with a differently named package. If a dependency requires rouge, install the fork alongside it and then remove the original's files:

pip install fitzyracing-rouge
pip uninstall -y rouge
pip install --force-reinstall --no-deps fitzyracing-rouge   # restore the files the uninstall removed

Afterwards pip check reports <package> requires rouge, which is not installed; that is expected. Repeat the steps if a later install pulls rouge back in.

uv users can do this properly with an override that drops the original:

# pyproject.toml
[project]
dependencies = ["fitzyracing-rouge", "...the package that depends on rouge..."]

[tool.uv]
override-dependencies = ["rouge; sys_platform == 'never'"]

(or uv pip install --override overrides.txt ... with that same line in overrides.txt).

Tests

pip install -e . pytest
pytest tests/test_fork_fixes.py

tests/test_fork_fixes.py fails on rouge 1.0.1 and passes on this fork; it also pins scores that must stay identical to 1.0.1.

Rouge

A full Python librarie for the ROUGE metric (paper).

Disclaimer

This implementation is independant from the "official" ROUGE script (aka. ROUGE-155).
Results may be slighlty different, see discussions in #2.

Quickstart

Clone & Install

git clone https://github.com/pltrdy/rouge
cd rouge
python setup.py install
# or
pip install -U .

or from pip:

pip install rouge

Use it from the shell (JSON Output)

$rouge -h
usage: rouge [-h] [-f] [-a] hypothesis reference

Rouge Metric Calculator

positional arguments:
  hypothesis  Text of file path
  reference   Text or file path

optional arguments:
  -h, --help  show this help message and exit
  -f, --file  File mode
  -a, --avg   Average mode

e.g.

# Single Sentence
rouge "transcript is a written version of each day 's cnn student" \
      "this page includes the show transcript use the transcript to help students with"

# Scoring using two files (line by line)
rouge -f ./tests/hyp.txt ./ref.txt

# Avg scoring - 2 files
rouge -f ./tests/hyp.txt ./ref.txt --avg

As a library

Score 1 sentence

from rouge import Rouge 

hypothesis = "the #### transcript is a written version of each day 's cnn student news program use this transcript to he    lp students with reading comprehension and vocabulary use the weekly newsquiz to test your knowledge of storie s you     saw on cnn student news"

reference = "this page includes the show transcript use the transcript to help students with reading comprehension and     vocabulary at the bottom of the page , comment for a chance to be mentioned on cnn student news . you must be a teac    her or a student age # # or older to request a mention on the cnn student news roll call . the weekly newsquiz tests     students ' knowledge of even ts in the news"

rouge = Rouge()
scores = rouge.get_scores(hypothesis, reference)

Output:

[
  {
    "rouge-1": {
      "f": 0.4786324739396596,
      "p": 0.6363636363636364,
      "r": 0.3835616438356164
    },
    "rouge-2": {
      "f": 0.2608695605353498,
      "p": 0.3488372093023256,
      "r": 0.20833333333333334
    },
    "rouge-l": {
      "f": 0.44705881864636676,
      "p": 0.5277777777777778,
      "r": 0.3877551020408163
    }
  }
]

Note: "f" stands for f1_score, "p" stands for precision, "r" stands for recall.

Score multiple sentences

import json
from rouge import Rouge

# Load some sentences
with open('./tests/data.json') as f:
  data = json.load(f)

hyps, refs = map(list, zip(*[[d['hyp'], d['ref']] for d in data]))
rouge = Rouge()
scores = rouge.get_scores(hyps, refs)
# or
scores = rouge.get_scores(hyps, refs, avg=True)

Output (avg=False): a list of n dicts:

{"rouge-1": {"f": _, "p": _, "r": _}, "rouge-2" : { .. }, "rouge-l": { ... }}

Output (avg=True): a single dict with average values:

{"rouge-1": {"f": _, "p": _, "r": _}, "rouge-2" : { ..     }, "rouge-l": { ... }}

Score two files (line by line)

Given two files hyp_path, ref_path, with the same number (n) of lines, calculate score for each of this lines, or, the average over the whole file.

from rouge import FilesRouge

files_rouge = FilesRouge()
scores = files_rouge.get_scores(hyp_path, ref_path)
# or
scores = files_rouge.get_scores(hyp_path, ref_path, avg=True)