Skip to content

DeepTeam is a framework to red team LLMs and AI agents.

BeaverTails framework added

A-Vamshiee38d3d9f622
diff --git a/deepteam/frameworks/__init__.py b/deepteam/frameworks/__init__.py
index 115a3933..a072cfa5 100644
--- a/deepteam/frameworks/__init__.py
+++ b/deepteam/frameworks/__init__.py
@@ -1,9 +1,11 @@
from .frameworks import AISafetyFramework
from .aegis.aegis import Aegis
from .owasp.owasp import OWASPTop10
+from .beavertails.beavertails import BeaverTails
__all__ = [
"AISafetyFramework",
"OWASPTop10",
"Aegis",
+ "BeaverTails"
]
diff --git a/deepteam/frameworks/aegis/aegis.py b/deepteam/frameworks/aegis/aegis.py
index 82a121e9..e18078d9 100644
--- a/deepteam/frameworks/aegis/aegis.py
+++ b/deepteam/frameworks/aegis/aegis.py
@@ -42,7 +42,7 @@ class Aegis(AISafetyFramework):
if len(attacks) < self.num_attacks:
raise ValueError(
- f"There are only {len(attacks)} harmful attacks that can be used in the Aegis dataset. Please set your 'num_attacks' to less than or equal to {len(attacks)}."
+ f"There are only {len(attacks)} harmful attacks that can be used in the {self.get_name()} dataset. Please set your 'num_attacks' to less than or equal to {len(attacks)}."
)
sampled = random.sample(attacks, k=self.num_attacks)
diff --git a/deepteam/frameworks/beavertails/beavertails.py b/deepteam/frameworks/beavertails/beavertails.py
new file mode 100644
index 00000000..5bd6c455
--- /dev/null
+++ b/deepteam/frameworks/beavertails/beavertails.py
@@ -0,0 +1,118 @@
+import asyncio
+import random
+from typing import Literal
+from deepeval.models import DeepEvalBaseLLM
+from tqdm import tqdm
+
+from deepteam.frameworks import AISafetyFramework
+from deepteam.attacks.multi_turn.types import CallbackType
+from deepteam.frameworks.beavertails.types import BeaverTailsType
+from deepteam.test_case import RTTestCase
+from deepteam.metrics import HarmMetric
+
+
+class BeaverTails(AISafetyFramework):
+ _DATASET = "PKU-Alignment/BeaverTails"
+
+ def __init__(
+ self,
+ num_attacks: int = 15,
+ data_split: Literal["330k_train", "30k_train"] = "30k_train",
+ async_mode: bool = True,
+ verbose_mode: bool = False,
+ evaluation_model: Optional[Union[str, DeepEvalBaseLLM]] = "gpt-4o",
+ ):
+ self.name = "BeaverTails"
+ self.description = "BeaverTails is an AI safety-focused collection comprising a series of datasets, it contains human-labeled data consisting of question-answer (QA) pairs from 14 different harm categories."
+ self.num_attacks = num_attacks
+ self.async_mode = async_mode
+ self.verbose_mode = verbose_mode
+ self.evaluation_model = evaluation_model
+ self.data_split = data_split
+ self._has_dataset = True
+
+ def load_dataset(self):
+ from datasets import load_dataset as hf_load_dataset
+
+ dataset = hf_load_dataset(self._DATASET, split=self.data_split)
+ attacks = [
+ row for row in dataset
+ if row["is_safe"] == False
+ ]
+
+ if len(attacks) < self.num_attacks:
+ raise ValueError(
+ f"There are only {len(attacks)} harmful attacks that can be used in the {self.get_name()} dataset. Please set your 'num_attacks' to less than or equal to {len(attacks)}."
+ )
+
+ sampled = random.sample(attacks, k=self.num_attacks)
+ test_cases = [
+ RTTestCase(
+ input=attack["prompt"],
+ vulnerability=" ".join([category for category, boolean in attack["category"].items() if boolean]),
+ vulnerability_type=BeaverTailsType.BEAVERTAILS,
+ )
+ for attack in sampled
+ ]
+ self.test_cases = test_cases
+ return self.test_cases
+
+ def assess(
+ self,
+ model_callback: CallbackType,
+ pbar: tqdm,
+ ignore_errors: bool = True,
+ ):
+ for test_case in self.test_cases:
+ try:
+ test_case.actual_output = model_callback(test_case.input)
+ metric = self._get_metric(test_case.vulnerability)
+ metric.measure(test_case)
+ test_case.score = metric.score
+ test_case.reason = metric.reason
+ except Exception as e:
+ if ignore_errors:
+ test_case.error = e.message
+ else:
+ raise
+ pbar.update(1)
+
+ return self.test_cases
+
+ async def a_assess(
+ self,
+ model_callback: CallbackType,
+ pbar: tqdm,
+ ignore_errors: bool = True,
+ ):
+
+ async def evaluate_test_case(test_case):
+ try:
+ test_case.actual_output = await model_callback(test_case.input)
+ metric = self._get_metric(test_case.vulnerability)
+ await metric.a_measure(test_case)
+ test_case.score = metric.score
+ test_case.reason = metric.reason
+ except Exception as e:
+ if ignore_errors:
+ test_case.error = e.message
+ else:
+ raise
+ pbar.update(1)
+
+ tasks = [evaluate_test_case(tc) for tc in self.test_cases]
+
+ await asyncio.gather(*tasks)
+
+ return self.test_cases
+
+ def _get_metric(self, harm_category: str):
+ return HarmMetric(
+ harm_category=harm_category,
+ model=self.evaluation_model,
+ async_mode=self.async_mode,
+ verbose_mode=self.verbose_mode,
+ )
+
+ def get_name(self):
+ return "BeaverTails"
diff --git a/deepteam/frameworks/beavertails/types.py b/deepteam/frameworks/beavertails/types.py
new file mode 100644
index 00000000..7758ae95
--- /dev/null
+++ b/deepteam/frameworks/beavertails/types.py
@@ -0,0 +1,6 @@
+from enum import Enum
+from typing import Literal
+
+
+class BeaverTailsType(Enum):
+ BEAVERTAILS = "beaver_tails_safety_dataset"