pretty_name: HarmProfile
license: apache-2.0
task_categories:

• text-generation
  tags:
• safety
• red-teaming
• synthetic
• harmful-content
  size_categories:
• 10K<n<100K
  configs:
• config_name: all_57
  default: true
  data_files:
  • split: train
    path: data/all_57/*.parquet
• config_name: high_risk_15
  data_files:
  • split: train
    path: data/high_risk_15/*.parquet

HarmProfile

HarmProfile is a structured safety and red-teaming dataset assembled from approved generation runs.

Content warning: This dataset contains synthetic prompts and responses involving harmful, illegal, abusive, explicit, self-harm, and other high-risk topics. Some categories may be especially sensitive. Use access controls and avoid rendering rows in logs, previews, notebooks, or monitoring systems unless necessary.

Load the dataset

from datasets import load_dataset

# Load the complete dataset or the high-risk subset.
all_57 = load_dataset(
    "Freshma/HarmProfile", "all_57", split="train", token=True
)
high_risk_15 = load_dataset(
    "Freshma/HarmProfile", "high_risk_15", split="train", token=True
)

# Load one category.
category = "cbrn"
category_dataset = load_dataset(
    "Freshma/HarmProfile",
    data_files=f"data/all_57/{category}.parquet",
    split="train",
    token=True,
)

Dataset structure

Both configurations use the train split. Each category is stored in one Zstandard-compressed Parquet file:

data/
├── all_57/          # 57 files, one per category
└── high_risk_15/    # 15 files, one per category

ALL_57 lists every category in the complete all_57 configuration. HIGH_RISK_15 lists the 15-category subset packaged as the high_risk_15 configuration. Each category name maps directly to data/all_57/<category>.parquet and can be passed to the loading pattern above.

HIGH_RISK_15 = [
    "cbrn",
    "copyright_reproduction",
    "csam",
    "dehumanization",
    "doxxing",
    "election_interference",
    "financial_scam",
    "harassment",
    "health_medical_misinfo",
    "human_trafficking",
    "jailbreak",
    "malware",
    "medical_advice",
    "suicide",
    "violence_incitement",
]

ALL_57 = [
    "academic_dishonesty",
    "adult_explicit",
    "animal_cruelty",
    "cbrn",
    "conspiracy_narrative",
    "copyright_reproduction",
    "critical_infrastructure",
    "csam",
    "defamation",
    "dehumanization",
    "document_forgery",
    "doxxing",
    "drugs",
    "eating_disorder",
    "election_interference",
    "evasion",
    "exploit_code",
    "explosives",
    "financial_advice",
    "financial_scam",
    "general_factual_misinfo",
    "guardrail_bypass",
    "harassment",
    "health_medical_misinfo",
    "human_trafficking",
    "illegal_firearms",
    "illegal_gambling",
    "impersonation",
    "jailbreak",
    "legal_advice",
    "malware",
    "market_manipulation",
    "medical_advice",
    "mental_health_crisis",
    "minor_grooming",
    "money_laundering",
    "non_consensual_sexual",
    "other_group_discrimination",
    "phishing_social_engineering",
    "pii_leak",
    "political_campaigning",
    "prompt_injection",
    "property_crime",
    "protected_attribute_hate",
    "science_denial",
    "self_injury",
    "spam",
    "spyware_surveillance_tool",
    "state_subversion_separatism",
    "suicide",
    "surveillance_stalking",
    "trade_secret",
    "trademark_misuse",
    "unauthorized_access",
    "violence_graphic",
    "violence_incitement",
    "weapons_trafficking",
]

Each record contains 8 columns:

Citation

If you use HarmProfile, please cite:

@misc{ma2026harmprofilecharacterizingharmfuldistributions,
  title         = {HarmProfile: Characterizing Harmful Distributions in Frontier LLMs},
  author        = {Zhouyuan Ma and Yutao Wu and Hanxun Huang and Xiang Zheng and Xiao Liu and Yixin Cao and Zuxuan Wu and Xingjun Ma and Yu-Gang Jiang},
  year          = {2026},
  eprint        = {2608.14577},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  url           = {https://arxiv.org/abs/2608.14577}
}