# Copyright (c) Microsoft Corporation. # Licensed under the MIT license. import json import logging import re from pyrit.datasets.seed_datasets.remote.remote_dataset_loader import ( _RemoteDatasetLoader, ) from pyrit.models import Modality, SeedDataset, SeedPrompt, SeedUnion from pyrit.models.harm_category import HarmCategory logger = logging.getLogger(__name__) class _SaladBenchDataset(_RemoteDatasetLoader): """ Loader for the SALAD-Bench dataset from HuggingFace. SALAD-Bench is a hierarchical and comprehensive safety benchmark for large language models. It organizes harmful questions into 6 domains, 16 tasks, and 65+ categories, totaling about 30k questions. It covers QA, multiple choice, attack-enhanced, and defense-enhanced variants. References: - https://huggingface.co/datasets/walledai/SaladBench - [@li2024saladbench] - https://github.com/OpenSafetyLab/SALAD-BENCH License: Apache 2.0 Warning: This dataset contains harmful and unsafe content designed for safety evaluation. """ HF_DATASET_NAME: str = "walledai/SaladBench" HARM_CATEGORY_ALIAS_OVERRIDES: dict[str, list[HarmCategory]] = { "academic dishonesty": [HarmCategory.PLAGIARISM], "adult content": [HarmCategory.SEXUAL_CONTENT], "age stereotyping": [HarmCategory.HATE_SPEECH, HarmCategory.REPRESENTATIONAL], "biological and chemical harms": [HarmCategory.CBRN], "campaign materials": [HarmCategory.CAMPAIGNING], "child abuse": [HarmCategory.GROOMING, HarmCategory.SEXUAL_CONTENT, HarmCategory.CHILD_LEAKAGE], "common sense": [HarmCategory.DANGEROUS_SITUATIONS], "copyright infringement": [HarmCategory.COPYRIGHT], "cyber attack": [HarmCategory.MALWARE, HarmCategory.COORDINATION_HARM], "defamation": [HarmCategory.REPUTATIONAL_DAMAGE], "defamation someone": [HarmCategory.REPUTATIONAL_DAMAGE], "disability stereotyping": [HarmCategory.HATE_SPEECH, HarmCategory.REPRESENTATIONAL], "disinformation": [HarmCategory.INFO_INTEGRITY], "dissemination of dangerous information": [HarmCategory.DANGEROUS_SITUATIONS], "drug-related crimes": [ HarmCategory.DRUG_USE, HarmCategory.REGULATED_GOODS, HarmCategory.COORDINATION_HARM, ], "enjoying someone else's pain": [HarmCategory.HARASSMENT], "environmental crimes": [HarmCategory.COORDINATION_HARM], "erosion of trust in public information": [HarmCategory.INFO_INTEGRITY], "erotic chat": [HarmCategory.SEXUAL_CONTENT, HarmCategory.SEXUAL_SOLICITATION], "fake news": [HarmCategory.INFO_INTEGRITY], "fake online content": [HarmCategory.DECEPTION, HarmCategory.INFO_INTEGRITY], "fake review generation": [HarmCategory.DECEPTION, HarmCategory.REPUTATIONAL_DAMAGE], "financial crimes": [HarmCategory.SCAMS, HarmCategory.COORDINATION_HARM], "financial data leaks": [HarmCategory.PPI, HarmCategory.PROPRIETARY_INFO], "financial fraud": [HarmCategory.SCAMS, HarmCategory.DECEPTION], "forge identity": [HarmCategory.IMPERSONATION, HarmCategory.FAKE_ACCOUNTS], "fraud or deceptive action": [HarmCategory.SCAMS, HarmCategory.DECEPTION], "gambling": [HarmCategory.GAMBLING], "gender stereotyping": [HarmCategory.HATE_SPEECH, HarmCategory.REPRESENTATIONAL], "government secrets": [HarmCategory.HIGH_RISK_GOVERNMENT, HarmCategory.PROPRIETARY_INFO], "group defamation": [ HarmCategory.HATE_SPEECH, HarmCategory.REPUTATIONAL_DAMAGE, HarmCategory.REPRESENTATIONAL, ], "harass, threaten, or bully an individual": [HarmCategory.HARASSMENT, HarmCategory.VIOLENT_THREATS], "hate speech": [HarmCategory.HATE_SPEECH], "historical revisionism": [HarmCategory.HISTORICAL_EVENTS_BIAS, HarmCategory.INFO_INTEGRITY], "human autonomy & integrity": [HarmCategory.COVERT_TARGETED, HarmCategory.DECEPTION], "illegal activities": [HarmCategory.COORDINATION_HARM], "illegal drugs and regulated/controlled substances": [HarmCategory.DRUG_USE, HarmCategory.REGULATED_GOODS], "illegal law advice": [HarmCategory.LEGAL_ADVICE, HarmCategory.COORDINATION_HARM], "illegitimate surveillance": [HarmCategory.COVERT_TARGETED, HarmCategory.PPI], "inferring personal information": [HarmCategory.PROTECTED_INFERENCE, HarmCategory.PPI], "influence operations": [HarmCategory.INFO_INTEGRITY, HarmCategory.COVERT_TARGETED], "information & safety": [HarmCategory.DANGEROUS_SITUATIONS], "insult": [HarmCategory.HARASSMENT, HarmCategory.PROFANITY], "labor issues": [HarmCategory.ALLOCATION], "law advice": [HarmCategory.LEGAL_ADVICE], "leaking private information": [HarmCategory.PPI], "location stereotyping": [HarmCategory.HATE_SPEECH, HarmCategory.REPRESENTATIONAL], "malicious use": [HarmCategory.COORDINATION_HARM], "malware generation": [HarmCategory.MALWARE], "management or operation of critical infrastructure": [HarmCategory.INFRASTRUCTURE_RISK], "manipulation of public opinion": [HarmCategory.INFO_INTEGRITY, HarmCategory.COVERT_TARGETED], "medical advice": [HarmCategory.HEALTH_DIAGNOSIS], "mimicking writing style": [HarmCategory.IMPERSONATION], "misinformation harms": [HarmCategory.INFO_INTEGRITY], "multi-level marketing": [HarmCategory.MLM], "paypal lending": [HarmCategory.LENDING], "persuasion and manipulation": [HarmCategory.COVERT_TARGETED, HarmCategory.DECEPTION], "plagiarism": [HarmCategory.PLAGIARISM], "pornography": [HarmCategory.SEXUAL_CONTENT], "privacy infringement": [HarmCategory.PPI], "propagating misconceptions/false beliefs": [HarmCategory.INFO_INTEGRITY], "pseudo-pharmaceuticals": [HarmCategory.PSEUDO_PHARMA], "psychological manipulation": [HarmCategory.EMOTIONAL, HarmCategory.COVERT_TARGETED], "racial and ethnic stereotyping": [HarmCategory.HATE_SPEECH, HarmCategory.REPRESENTATIONAL], "radicalization": [HarmCategory.VIOLENT_EXTREMISM], "religious stereotyping": [HarmCategory.HATE_SPEECH, HarmCategory.REPRESENTATIONAL], "representation & toxicity": [HarmCategory.HATE_SPEECH, HarmCategory.REPRESENTATIONAL], "risky financial practices": [HarmCategory.FINANCIAL_ADVICE, HarmCategory.FINANCIAL_ELIGIBILITY], "scams": [HarmCategory.SCAMS], "science denialism": [HarmCategory.INFO_INTEGRITY, HarmCategory.PUBLIC_HEALTH], "security threats": [HarmCategory.DANGEROUS_SITUATIONS, HarmCategory.INFRASTRUCTURE_RISK], "self-harm": [HarmCategory.SELF_HARM], "sexual activity": [HarmCategory.SEXUAL_CONTENT, HarmCategory.SEXUAL_SOLICITATION], "sexual offenses": [ HarmCategory.SEXUAL_CONTENT, HarmCategory.SEXUAL_SOLICITATION, HarmCategory.NONCONSENSUAL_UPLOAD, ], "social media rumors": [HarmCategory.INFO_INTEGRITY], "socioeconomic harms": [HarmCategory.ALLOCATION], "spam": [HarmCategory.SPAM], "terrorism": [HarmCategory.VIOLENT_EXTREMISM, HarmCategory.COORDINATION_HARM], "theft": [HarmCategory.COORDINATION_HARM], "toxic content": [HarmCategory.HARASSMENT, HarmCategory.HATE_SPEECH, HarmCategory.PROFANITY], "trade affairs": [HarmCategory.ALLOCATION], "trade and compliance": [HarmCategory.COORDINATION_HARM], "traffic and driving offenses": [HarmCategory.COORDINATION_HARM], "unfair representation": [HarmCategory.REPRESENTATIONAL, HarmCategory.ALLOCATION], "violation of personal property": [HarmCategory.COORDINATION_HARM], "violent content": [HarmCategory.VIOLENT_CONTENT], "violent crimes": [ HarmCategory.VIOLENT_CONTENT, HarmCategory.VIOLENT_THREATS, HarmCategory.COORDINATION_HARM, ], "weapon generation and transportation": [HarmCategory.REGULATED_GOODS], } # Metadata modalities: tuple[Modality, ...] = (Modality.TEXT,) size: str = "huge" # 21318 harmful questions across 6 domains, 16 tasks, 65+ categories tags: frozenset[str] = frozenset({"default", "safety", "jailbreak"}) def __init__( self, *, config: str = "prompts", split: str = "base", ) -> None: """ Initialize the SALAD-Bench dataset loader. Args: config: Dataset configuration. Defaults to "prompts". split: Dataset split to load. One of "base", "attackEnhanced", "defenseEnhanced". Defaults to "base". """ self.config = config self.split = split @property def dataset_name(self) -> str: """The dataset name.""" return "salad_bench" @staticmethod def _parse_category(category: str) -> str: """ Strip leading identifier like 'O6: ' from a category string. Args: category (str): The category string to parse. Returns: str: The category string without the leading identifier. """ return re.sub(r"^O\d+:\s*", "", category) async def fetch_dataset_async(self, *, cache: bool = True) -> SeedDataset: """ Fetch SALAD-Bench dataset from HuggingFace and return as SeedDataset. Args: cache: Whether to cache the fetched dataset. Defaults to True. Returns: SeedDataset: A SeedDataset containing the SALAD-Bench prompts. """ logger.info(f"Loading SALAD-Bench dataset from {self.HF_DATASET_NAME}") data = await self._fetch_from_huggingface_async( dataset_name=self.HF_DATASET_NAME, config=self.config, split=self.split, cache=cache, ) authors = [ "Lijun Li", "Bowen Dong", "Ruohui Wang", "Xuhao Hu", "Wangmeng Zuo", "Dahua Lin", "Yu Qiao", "Jing Shao", ] description = ( "SALAD-Bench is a hierarchical and comprehensive safety benchmark for large language " "models (ACL 2024). It contains about 30k questions organized into 6 domains, 16 tasks, " "and 65+ categories, with base, attack-enhanced, and defense-enhanced variants." ) source_url = f"https://huggingface.co/datasets/{self.HF_DATASET_NAME}" groups = [ "Shanghai Artificial Intelligence Laboratory", "Harbin Institute of Technology", "Beijing Institute of Technology", "The Chinese University of Hong Kong", "The Hong Kong Polytechnic University", ] seed_prompts: list[SeedUnion] = [] for item in data: parsed_categories = [self._parse_category(c) for c in item["categories"]] metadata: dict[str, str | int] = {"categories": json.dumps(item["categories"])} if source := item.get("source"): metadata["original_source"] = source seed_prompts.append( SeedPrompt( value=item["prompt"], data_type="text", dataset_name=self.dataset_name, harm_categories=self._standardize_harm_categories( parsed_categories, alias_overrides=self.HARM_CATEGORY_ALIAS_OVERRIDES, ), description=description, source=source_url, authors=authors, groups=groups, metadata=metadata, ) ) logger.info(f"Successfully loaded {len(seed_prompts)} prompts from SALAD-Bench dataset") return SeedDataset(seeds=seed_prompts, dataset_name=self.dataset_name)