| {"safe": 0.5, "mild_toxicity": 0.7, "severe_toxicity": 0.4, "harassment": 0.5, "hate_speech": 0.45, "threat": 0.4, "insult": 0.55, "profanity": 0.6, "sexual_content": 0.45, "violence": 0.4, "self_harm": 0.35, "extremism": 0.4, "illegal_instruction": 0.45, "jailbreak_attempt": 0.45, "prompt_injection": 0.45, "obfuscated_toxicity": 0.5, "misinformation_risk": 0.5} |