feat: add ROUGE score descriptor and dataset metric (#1318)

[mukund1985]

Summary

Closes #1318

Implements ROUGE (Recall-Oriented Understudy for Gisting Evaluation) score as a row-level Descriptor and a dataset-level Metric, as requested in the issue.

Changes

RougeScore descriptor — row-level

File: src/evidently/descriptors/_rouge_score.py

Computes ROUGE score between a prediction column and a reference column for each row. Returns a numeric value in [0, 1].

from evidently.descriptors import RougeScore

dataset = Dataset.from_pandas(df, data_definition=DataDefinition(), descriptors=[
    RougeScore("response", "ground_truth", rouge_type="rouge1", alias="ROUGE-1 F1"),
    RougeScore("response", "ground_truth", rouge_type="rouge2", score_type="recall"),
    RougeScore("response", "ground_truth", rouge_type="rougeL"),
])

RougeScoreMetric — dataset-level aggregate

File: src/evidently/metrics/text_evals.py

A dedicated SingleValueMetric (not ColumnSummaryMetric) that computes the mean ROUGE score across all rows. Renders a counter widget + histogram distribution. Supports current-only and current+reference comparisons.

from evidently import Report
from evidently.metrics import RougeScoreMetric

report = Report([
    RougeScoreMetric(
        prediction_column="response",
        reference_column="ground_truth",
        rouge_type="rouge1",
    )
])
result = report.run(current_dataset, reference_dataset)

RougeScoreFeature — legacy feature layer

File: src/evidently/legacy/features/rouge_score_feature.py

Follows the same pattern as SemanticSimilarityFeature. Registered in _registry.py.

Parameters

Both RougeScore and RougeScoreMetric expose:

Parameter	Values	Default
rouge_type	rouge1, rouge2, rougeL, rougeLsum	rouge1
score_type	f (F1), precision, recall	f

Dependency

Added rouge-score>=0.1.2 to the [llm] optional extra — same group as sentence-transformers and transformers. Import is lazy (inside function bodies), so it only fails if the feature is used without pip install evidently[llm].

Added rouge_score.* to [[tool.mypy.overrides]] ignore list.

Tests

31 tests across 3 files, all passing:

• tests/features/test_rouge_score_feature.py — 13 tests (parametrised variants, NaN handling, invalid params, feature name)
• tests/descriptors/test_rouge_score_descriptor.py — 11 tests (all ROUGE types, precision/recall, alias, column listing, validation)
• tests/metrics/test_rouge_score_metric.py — 7 tests (score range, with/without reference, identical texts, all variants)

[mukund1985]

This is my first feature PR here so apologies if I've missed anything obvious. I tried to follow the same patterns as SemanticSimilarity and TextLength when putting it together. 31 tests pass locally. Also happy to split it into smaller PRs if that makes reviewing easier — just say the word.

[mukund1985]

@Liraim — would appreciate a review when you get a chance. Tests pass locally, happy to make any changes needed.

[mukund1985]

@DimaAmega — looks like CI hasn't triggered yet, could you approve the workflow run when you get a chance?