metadata
title: VoxSafeBench
tags:
- audio
- safety
- fairness
- privacy
configs:
- config_name: Safety-tier1
data_files:
- split: No_jailbreak
path: Safety-tier1/No_jailbreak/metadata.jsonl
- split: Singleturn_jailbreak
path: Safety-tier1/Singleturn_jailbreak/metadata.jsonl
- split: Multiturn_jailbreak
path: Safety-tier1/Multiturn_jailbreak/metadata.jsonl
- split: Agentic_Action_Risks
path: Safety-tier1/Agentic_Action_Risks/metadata.jsonl
- config_name: Safety-tier2
data_files:
- split: Child_voice
path: Safety-tier2/Child_voice/metadata.jsonl
- split: Emotion
path: Safety-tier2/Emotion/metadata.jsonl
- split: Impaired_capacity
path: Safety-tier2/Impaired_capacity/metadata.jsonl
- split: Child_presence
path: Safety-tier2/Child_presence/metadata.jsonl
- split: Symbolic_background
path: Safety-tier2/Symbolic_background/metadata.jsonl
- split: Unsafe_ambient
path: Safety-tier2/Unsafe_ambient/metadata.jsonl
- split: Overlap_instruction_injection
path: Safety-tier2/Overlap_instruction_injection/metadata.jsonl
- config_name: Fairness-tier1
data_files:
- split: test
path: Fairness-tier1/test/metadata.jsonl
- config_name: Fairness-tier2
data_files:
- split: test
path: Fairness-tier2/test/metadata.jsonl
- split: Bias_analysis
path: Fairness-tier2/Bias_analysis/metadata.jsonl
- config_name: Privacy-tier1
data_files:
- split: Hard_privacy
path: Privacy-tier1/Hard_privacy/metadata.jsonl
- split: Soft_privacy
path: Privacy-tier1/Soft_privacy/metadata.jsonl
- config_name: Privacy-tier2
data_files:
- split: Audio_conditioned_privacy
path: Privacy-tier2/Audio_conditioned_privacy/metadata.jsonl
- split: Interactional_privacy
path: Privacy-tier2/Interactional_privacy/metadata.jsonl
dataset_info:
features:
- name: system_prompt
dtype: string
- name: clean_audio_file_name
dtype: audio
- name: diverse_audio_file_name
dtype: audio
- name: audio_file_name
dtype: audio
- name: transcript
dtype: string
- name: source_text
dtype: string
- name: super_category
dtype: string
- name: jailbreak_type
dtype: string
- name: task_type
dtype: string
- name: language
dtype: string
- name: query
dtype: string
- name: harmless_text
dtype: string
- name: harm_text
dtype: string
- name: dimension
dtype: string
- name: category
dtype: string
- name: flipped_audio_file_name
dtype: audio
- name: note
dtype: string
- name: Language
dtype: string
- name: turn1_audio_file_name
dtype: audio
- name: turn1_transcript
dtype: string
- name: turn2_audio_file_name
dtype: audio
- name: turn3_audio_file_name
dtype: audio
- name: turn3_transcript
dtype: string
- name: GT
dtype: string
- name: paralinguistic
dtype: string
- name: acoustic
dtype: string
license: apache-2.0
language:
- en
- zh
size_categories:
- 10K<n<100K
VoxSafeBench
This dataset is uploaded as raw files (JSONL + audio), not parquet.
Subset: Safety-tier1
Split: No_jailbreak (8708 samples)
Columns: system_prompt, clean_audio_file_name, diverse_audio_file_name, transcript, super_category, task_type, language, query
Split: Singleturn_jailbreak (2516 samples)
Columns: system_prompt, audio_file_name, transcript, source_text, super_category, jailbreak_type, task_type, language, query
Split: Multiturn_jailbreak (240 samples)
Columns: system_prompt, turn1_audio_file_name, turn1_transcript, turn2_audio_file_name, turn2_transcript, turn3_audio_file_name, turn3_transcript, source_text, super_category, language
Split: Safety-tier2/Child_voice
Total samples: 400
Split: Safety-tier2/Emotion
Total samples: 160
Split: Safety-tier2/Impaired_capacity
Total samples: 293
Split: Safety-tier2/Child_presence
Total samples: 400
Split: Safety-tier2/Symbolic_background
Total samples: 320
Split: Safety-tier2/Unsafe_ambient
Total samples: 400
Split: Safety-tier2/Overlap_instruction_injection
Total samples: 552
Split: Fairness-tier1/test
Total samples: 2312
Split: Fairness-tier2/test
Total samples: 2050
Split: Fairness-tier2/Bias_analysis
Total samples: 2218
Split: Safety-tier1/Agentic_Action_Risks
Total samples: 873
Split: Privacy-tier1/Hard_privacy
Total samples: 800
Split: Privacy-tier1/Soft_privacy
Total samples: 800
Split: Privacy-tier2/Audio_conditioned_privacy
Total samples: 400
Split: Privacy-tier2/Interactional_privacy
Total samples: 586