chore: revive semantic similarity test

src/nemo_safe_synthesizer/evaluation/components/text_semantic_similarity.py

@@ -5,14 +5,14 @@
 
 import logging
 from functools import cached_property
+from typing import TYPE_CHECKING
 
 import numpy as np
 import numpy.typing as npt
 import pandas as pd
 from numpy.linalg import norm
 from pydantic import BaseModel, ConfigDict, Field
 from scipy.stats import ks_2samp
-from sentence_transformers import SentenceTransformer
 from tenacity import (
     RetryError,
     Retrying,
@@ -35,6 +35,9 @@
 from ...observability import get_logger
 from . import multi_modal_figures as figures
 
+if TYPE_CHECKING:
+    from sentence_transformers import SentenceTransformer
+
 logger = get_logger(__name__)
 
 
@@ -237,6 +240,11 @@ def _preprocess_text_data(text_data: pd.Series, nrows: int) -> pd.Series:
     @staticmethod
     def _init_sentence_transformer_model() -> SentenceTransformer | None:
         """Load the sentence transformer model with exponential-backoff retries."""
+        try:
+            from sentence_transformers import SentenceTransformer
+        except ImportError:
+            return None
+
         try:
             for attempt in Retrying(
                 # TODO(PLAT-2537): Temporarily increase retries, but we will bundle this model next, so download is not required.

tests/e2e/test_safe_synthesizer.py

@@ -40,6 +40,12 @@
 logger = get_logger(__name__)
 
 
+def _assert_evaluation_report_rendered(report_html: str | None) -> None:
+    assert report_html is not None
+    assert "Synthetic Quality Score" in report_html
+    assert "Text Semantic Similarity" in report_html
+
+
 @pytest.mark.e2e
 @pytest.mark.requires_gpu
 @pytest.mark.timeout(1000)
@@ -66,6 +72,7 @@ def test_train_and_generate_dp(fixture_financial_transactions_dataset, fixture_s
     assert result.summary.timing.training_time_sec is not None and result.summary.timing.training_time_sec > 0
     assert result.summary.timing.generation_time_sec is not None and result.summary.timing.generation_time_sec > 0
     assert result.summary.timing.evaluation_time_sec is not None and result.summary.timing.evaluation_time_sec > 0
+    _assert_evaluation_report_rendered(result.evaluation_report_html)
 
 
 @pytest.mark.e2e
@@ -89,3 +96,4 @@ def test_train_and_generate_defaults(fixture_financial_transactions_dataset, fix
     assert result.summary.timing.training_time_sec is not None and result.summary.timing.training_time_sec > 0
     assert result.summary.timing.generation_time_sec is not None and result.summary.timing.generation_time_sec > 0
     assert result.summary.timing.evaluation_time_sec is not None and result.summary.timing.evaluation_time_sec > 0
+    _assert_evaluation_report_rendered(result.evaluation_report_html)

tests/evaluation/reports/test_multimodal_report.py

@@ -1,22 +1,36 @@
 # SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 # SPDX-License-Identifier: Apache-2.0
 
+import json
+
 import pandas as pd
 import pytest
 
-# Skip all tests in this module if sentence_transformers is not available
-pytest.importorskip(
-    "sentence_transformers",
-    reason="sentence_transformers is required for these tests (install with: uv sync --extra cpu)",
-)
-
 from nemo_safe_synthesizer.config.evaluate import EvaluationParameters
 from nemo_safe_synthesizer.config.parameters import SafeSynthesizerParameters
+from nemo_safe_synthesizer.evaluation.components.text_semantic_similarity import TextSemanticSimilarity
 from nemo_safe_synthesizer.evaluation.data_model.evaluation_datasets import EvaluationDatasets
-from nemo_safe_synthesizer.evaluation.data_model.evaluation_score import Grade
+from nemo_safe_synthesizer.evaluation.data_model.evaluation_score import EvaluationScore, Grade
+from nemo_safe_synthesizer.evaluation.reports.multimodal import multimodal_report as multimodal_report_module
 from nemo_safe_synthesizer.evaluation.reports.multimodal.multimodal_report import MultimodalReport
 
 
+@pytest.fixture(autouse=True)
+def stub_text_semantic_similarity(monkeypatch: pytest.MonkeyPatch) -> None:
+    """Keep report assembly tests independent of sentence-transformer inference."""
+
+    class StubTextSemanticSimilarity:
+        @staticmethod
+        def from_evaluation_datasets(*_args, **_kwargs) -> TextSemanticSimilarity:
+            return TextSemanticSimilarity(score=EvaluationScore.finalize_grade(9.0, 9.0))
+
+    monkeypatch.setattr(
+        multimodal_report_module,
+        "TextSemanticSimilarity",
+        StubTextSemanticSimilarity,
+    )
+
+
 def _minimal_multimodal_report() -> MultimodalReport:
     training_df = pd.DataFrame({"x": [1, 2], "y": [3, 4]})
     synthetic_df = pd.DataFrame({"x": [1, 2], "y": [3, 4]})
@@ -38,7 +52,6 @@ def test_jinja_context_job_id_set_when_nemo_job_id_present(monkeypatch: pytest.M
     assert ctx["job_id"] == "cluster-job-abc123"
 
 
-@pytest.mark.slow
 def test_from_dataframes_applies_sqs_report_config(fixture_training_df, fixture_synthetic_df, fixture_test_df) -> None:
     """``sqs_report_rows`` / ``sqs_report_columns`` from config drive the actual subsampling.
 
@@ -72,13 +85,12 @@ def test_from_dataframes_applies_sqs_report_config(fixture_training_df, fixture_
     assert report.evaluation_datasets.synthetic_cols == target_cols
 
 
-@pytest.mark.slow
 def test_multimodal_report(
-    fixture_training_df_5k, fixture_synthetic_df_5k, fixture_test_df, fixture_skip_privacy_metrics_config
+    fixture_training_df, fixture_synthetic_df, fixture_test_df, fixture_skip_privacy_metrics_config
 ):
     report = MultimodalReport.from_dataframes(
-        training=fixture_training_df_5k,
-        synthetic=fixture_synthetic_df_5k,
+        training=fixture_training_df,
+        synthetic=fixture_synthetic_df,
         test=fixture_test_df,
         config=fixture_skip_privacy_metrics_config,
     )
@@ -89,15 +101,15 @@ def test_multimodal_report(
 
     report_dict = report.get_dict()
     assert len(report_dict) == 6
-    assert report_dict["Synthetic Quality Score"] == {
-        "raw_score": 9.7935,
+    assert report_dict["Text Semantic Similarity"] == {
+        "raw_score": 9.0,
         "grade": "Excellent",
-        "score": 9.8,
+        "score": 9.0,
         "notes": None,
     }
+    assert report_dict["Synthetic Quality Score"]["grade"] == "Excellent"
+    assert report_dict["Synthetic Quality Score"]["score"] > 0
 
-    report_json = report.get_json()
-    assert (
-        '"Synthetic Quality Score": {"raw_score": 9.7935, "grade": "Excellent", "score": 9.8, "notes": null}}'
-        in report_json
-    )
+    report_json = json.loads(report.get_json())
+    assert report_json["Text Semantic Similarity"] == report_dict["Text Semantic Similarity"]
+    assert report_json["Synthetic Quality Score"] == report_dict["Synthetic Quality Score"]