From 61be89cbbe85e0017cb58f10396c5ad13e403eff Mon Sep 17 00:00:00 2001
From: "zhao.wang" <57819425+Excelius-Wang@users.noreply.github.com>
Date: Thu, 10 Sep 2026 00:59:07 +0800
Subject: [PATCH] fix(metrics): handle period-only text in ROUGE evaluation

---
 swift/metrics/nlg.py            |  8 ++++++-
 tests/utils/test_nlg_metrics.py | 41 +++++++++++++++++++++++++++++++++
 2 files changed, 48 insertions(+), 1 deletion(-)

diff --git a/swift/metrics/nlg.py b/swift/metrics/nlg.py
index 0a17615eeb..53fcd6c564 100644
--- a/swift/metrics/nlg.py
+++ b/swift/metrics/nlg.py
@@ -25,7 +25,13 @@ def compute_rouge_bleu(preds: List[str], labels: List[str]):
                 metric.update(0.)
             continue
         rouge = Rouge()
-        scores = rouge.get_scores(' '.join(hypothesis), ' '.join(reference))[0]
+        hypothesis_text = ' '.join(hypothesis)
+        reference_text = ' '.join(reference)
+        # Rouge splits on periods and drops empty segments; a period-only string leaves no sentences.
+        if hypothesis_text.strip('.') and reference_text.strip('.'):
+            scores = rouge.get_scores(hypothesis_text, reference_text)[0]
+        else:
+            scores = {key: {'f': 0.0} for key in rouge.metrics}
         for k, v in scores.items():
             score_dict[k].update(v['f'])
         bleu_score = sentence_bleu([reference], hypothesis, smoothing_function=SmoothingFunction().method3)
diff --git a/tests/utils/test_nlg_metrics.py b/tests/utils/test_nlg_metrics.py
index 06667e2286..d5a3ae7fdd 100644
--- a/tests/utils/test_nlg_metrics.py
+++ b/tests/utils/test_nlg_metrics.py
@@ -30,6 +30,47 @@ def test_empty_prediction_counts_towards_mean(self):
             'bleu-4': 50.0,
         })
 
+    def test_period_only_prediction_counts_towards_mean(self):
+        exact_match = 'the cat is here'
+        for prediction in ['.', '...', '. .']:
+            with self.subTest(prediction=prediction):
+                scores = compute_rouge_bleu([exact_match, prediction], [exact_match, 'a different reference'])
+                self.assertEqual(scores, {'rouge-1': 50.0, 'rouge-2': 50.0, 'rouge-l': 50.0, 'bleu-4': 50.0})
+
+    def test_period_only_reference(self):
+        scores = compute_rouge_bleu(['the cat is here'], ['.'])
+        self.assertEqual(scores, {'rouge-1': 0.0, 'rouge-2': 0.0, 'rouge-l': 0.0, 'bleu-4': 0.0})
+
+    def test_bleu_keeps_period_tokens(self):
+        from nltk.translate.bleu_score import SmoothingFunction, sentence_bleu
+        scores = compute_rouge_bleu(['.'], ['.'])
+        expected_bleu = sentence_bleu([['.']], ['.'], smoothing_function=SmoothingFunction().method3)
+        self.assertEqual(scores['bleu-4'], round(expected_bleu * 100, 6))
+        self.assertGreater(scores['bleu-4'], 0)
+        for key in ['rouge-1', 'rouge-2', 'rouge-l']:
+            self.assertEqual(scores[key], 0)
+
+    def test_ordinary_punctuation_keeps_rouge_scores(self):
+        from rouge.rouge import Rouge
+        text = 'the cat is here.'
+        scores = compute_rouge_bleu([text], [text])
+        reference = Rouge().get_scores(text, text)[0]
+        for key, value in reference.items():
+            self.assertEqual(scores[key], round(value['f'] * 100, 6))
+
+    def test_spaced_periods_keep_existing_rouge_behavior(self):
+        from rouge.rouge import Rouge
+        text = '. .'
+        scores = compute_rouge_bleu([text], [text])
+        reference = Rouge().get_scores(text, text)[0]
+        for key, value in reference.items():
+            self.assertEqual(scores[key], round(value['f'] * 100, 6))
+
+    def test_unrelated_rouge_errors_are_not_suppressed(self):
+        with patch('rouge.rouge.Rouge.get_scores', side_effect=ValueError('unrelated failure')):
+            with self.assertRaisesRegex(ValueError, 'unrelated failure'):
+                compute_rouge_bleu(['the cat is here'], ['the cat is here'])
+
 
 if __name__ == '__main__':
     unittest.main()