Skip to content

Fully automatic censorship removal for language models

fix: escape prompts and responses printed with `print_responses` (#482)

Jay Hemnani17f8dff09fdd
diff --git a/src/heretic/scorers/keyword_rate.py b/src/heretic/scorers/keyword_rate.py
index 106f9fa..9895f03 100644
--- a/src/heretic/scorers/keyword_rate.py
+++ b/src/heretic/scorers/keyword_rate.py
@@ -2,6 +2,7 @@
# Copyright (C) 2025-2026 Philipp Emanuel Weidmann <[email protected]> + contributors
from pydantic import BaseModel, Field
+from rich.markup import escape
from heretic.config import DatasetSpecification, SingleDatasetSpecification
from heretic.scorer import Context, Score, Scorer
@@ -105,10 +106,12 @@ class KeywordRate(Scorer):
if self.settings.print_responses:
print()
- print(f"[bold]System prompt:[/] {prompt.system}")
- print(f"[bold]Prompt:[/] {prompt.user}")
+ print(f"[bold]System prompt:[/] {escape(prompt.system)}")
+ print(f"[bold]Prompt:[/] {escape(prompt.user)}")
if not response.strip():
response = "[italic]\\[empty][/]"
+ else:
+ response = escape(response)
print(
f"[bold]Response:[/] [{'red' if is_match else 'green'}]{response}[/]"
)
diff --git a/tests/test_keyword_rate.py b/tests/test_keyword_rate.py
new file mode 100644
index 0000000..2ef98c7
--- /dev/null
+++ b/tests/test_keyword_rate.py
@@ -0,0 +1,61 @@
+# SPDX-License-Identifier: AGPL-3.0-or-later
+# Copyright (C) 2025-2026 Philipp Emanuel Weidmann <[email protected]> + contributors
+
+import io
+import unittest
+from typing import cast
+from unittest.mock import patch
+
+from rich.console import Console
+
+from heretic.config import Settings as HereticSettings
+from heretic.scorer import Context
+from heretic.scorers.keyword_rate import KeywordRate, Settings
+from heretic.utils import Prompt
+
+
+class FakeContext:
+ def __init__(self, responses: list[str]) -> None:
+ self.responses = responses
+
+ def get_responses(self, prompts: list[Prompt]) -> list[str]:
+ return self.responses
+
+
+class KeywordRateTests(unittest.TestCase):
+ def test_print_responses_shows_brackets_literally(self) -> None:
+ scorer = KeywordRate(
+ heretic_settings=HereticSettings.model_construct(),
+ settings=Settings(print_responses=True),
+ )
+ scorer.prompts = [
+ Prompt(
+ system="You are a [helpful] assistant.",
+ user="Name a common misconception about [bird migration].",
+ ),
+ Prompt(
+ system="You are a helpful assistant.",
+ user="What does the [/INST] token mean?",
+ ),
+ ]
+ ctx = FakeContext(["Use arr[i] here.", "It ends the [/INST] block."])
+
+ output = io.StringIO()
+ console = Console(file=output, highlight=False, width=200)
+
+ with patch("heretic.scorers.keyword_rate.print", console.print):
+ score = scorer.get_score(cast(Context, ctx))
+
+ self.assertEqual(score.md_display, "0/2")
+ for text in [
+ "You are a [helpful] assistant.",
+ "[bird migration]",
+ "What does the [/INST] token mean?",
+ "Use arr[i] here.",
+ "It ends the [/INST] block.",
+ ]:
+ self.assertIn(text, output.getvalue())
+
+
+if __name__ == "__main__":
+ unittest.main()