FIX: Preserve text columns when loading human-labeled CSVs

pyrit/score/scorer_evaluation/human_labeled_dataset.py

@@ -15,6 +15,8 @@
 from pyrit.score.scorer_evaluation.metrics_type import MetricsType
 
 if TYPE_CHECKING:
+    from collections.abc import Hashable
+
     from pyrit.models.harm_definition import HarmDefinition
 
 logger = logging.getLogger(__name__)
@@ -271,11 +273,21 @@ def from_csv(
         if not harm_definition_version and parsed_harm_definition_version:
             harm_definition_version = parsed_harm_definition_version
 
+        # Preserve text before pandas can infer numeric or boolean values.
+        text_dtypes: dict[Hashable, type[str]] = dict.fromkeys(
+            (
+                cls.STANDARD_ASSISTANT_RESPONSE_COL,
+                cls.STANDARD_OBJECTIVE_COL,
+                cls.STANDARD_HARM_COL,
+                cls.STANDARD_DATA_TYPE_COL,
+            ),
+            str,
+        )
         # Try UTF-8 first, fall back to latin-1 for files with special characters
         try:
-            eval_df = pd.read_csv(csv_path, comment="#", encoding="utf-8")
+            eval_df = pd.read_csv(csv_path, comment="#", encoding="utf-8", dtype=text_dtypes)
         except UnicodeDecodeError:
-            eval_df = pd.read_csv(csv_path, comment="#", encoding="latin-1")
+            eval_df = pd.read_csv(csv_path, comment="#", encoding="latin-1", dtype=text_dtypes)
 
         # Drop rows where every column is NaN (e.g. trailing blank lines in the CSV)
         eval_df = eval_df.dropna(how="all")

tests/unit/score/test_human_labeled_dataset.py

@@ -1,6 +1,8 @@
 # Copyright (c) Microsoft Corporation.
 # Licensed under the MIT license.
 
+from pathlib import Path
+
 import pytest
 
 from pyrit.common.path import SCORER_EVALS_PATH
@@ -13,6 +15,35 @@
 )
 
 
+@pytest.mark.parametrize("value", ["00123", "1e3", "1.2300", "True"])
+@pytest.mark.parametrize("encoding", ["utf-8", "latin-1"])
+@pytest.mark.parametrize("metrics_type", [MetricsType.OBJECTIVE, MetricsType.HARM])
+def test_from_csv_preserves_numeric_and_boolean_text(
+    tmp_path: Path, value: str, encoding: str, metrics_type: MetricsType
+) -> None:
+    column = "objective" if metrics_type == MetricsType.OBJECTIVE else "harm_category"
+    label = "1" if metrics_type == MetricsType.OBJECTIVE else "0.5"
+    csv_path = tmp_path / "text.csv"
+    csv_path.write_text(
+        f"assistant_response,{column},human_score,notes\n{value},00456,{label},café\n",
+        encoding=encoding,
+    )
+
+    dataset = HumanLabeledDataset.from_csv(csv_path=csv_path, metrics_type=metrics_type, version="1.0")
+    entry = dataset.entries[0]
+    piece = entry.conversation[0].message_pieces[0]
+
+    assert piece.original_value == value
+    assert piece.converted_value == value
+    if isinstance(entry, ObjectiveHumanLabeledEntry):
+        assert entry.objective == "00456"
+        assert entry.human_scores == [True]
+    else:
+        assert isinstance(entry, HarmHumanLabeledEntry)
+        assert entry.harm_category == "00456"
+        assert entry.human_scores == [0.5]
+
+
 @pytest.fixture
 def sample_messages():
     return [