Skip to content

armgit / txtai

Public Imported from GitHub
0

💡 All-in-one AI framework for semantic search, LLM orchestration and language model workflows

Respect numeric zero label flattening thresholds (#1398)

HuaTNA224ac2de1aee
diff --git a/src/python/txtai/pipeline/text/labels.py b/src/python/txtai/pipeline/text/labels.py
index 5afd49ea..d2e2d27b 100644
--- a/src/python/txtai/pipeline/text/labels.py
+++ b/src/python/txtai/pipeline/text/labels.py
@@ -88,13 +88,13 @@ class Labels(HFPipeline):
for result in results:
if self.dynamic:
- if flatten:
+ if flatten is not None and flatten is not False:
result = [label for x, label in enumerate(result["labels"]) if result["scores"][x] >= threshold]
yield result[:1] if isinstance(flatten, bool) else result
else:
yield [(labels.index(label), result["scores"][x]) for x, label in enumerate(result["labels"])]
else:
- if flatten:
+ if flatten is not None and flatten is not False:
result = [x["label"] for x in result if x["score"] >= threshold and (not labels or x["label"] in labels)]
yield result[:1] if isinstance(flatten, bool) else result
else:
diff --git a/test/python/testpipeline/testtext/testlabels.py b/test/python/testpipeline/testtext/testlabels.py
index 6b050574..afb9ddae 100644
--- a/test/python/testpipeline/testtext/testlabels.py
+++ b/test/python/testpipeline/testtext/testlabels.py
@@ -59,6 +59,20 @@ class TestLabels(unittest.TestCase):
results = [l[0] for l in self.labels(["This is the best sentence ever", "This is terrible"], ["positive", "negative"], flatten=True)]
self.assertEqual(results, ["positive", "negative"])
+ def testLabelZeroThreshold(self):
+ """
+ Test numeric zero flattening thresholds with dynamic and fixed labels
+ """
+
+ text = "This is the best sentence ever"
+ for classifier, candidates in ((self.labels, ["positive", "negative"]), (Labels(dynamic=False), ["POSITIVE", "NEGATIVE"])):
+ with self.subTest(dynamic=classifier.dynamic):
+ self.assertCountEqual(classifier(text, candidates, flatten=0.0), candidates)
+ self.assertCountEqual(classifier([text], candidates, flatten=0.0)[0], candidates)
+
+ # Boolean False still requests the original score tuples.
+ self.assertEqual(classifier(text, candidates, flatten=False), classifier(text, candidates))
+
def testLabelFixed(self):
"""
Test labels with a fixed label text classification model