Get the app
misc 0 0 0

knoveleng/redbench

published by @knoveleng

Alignment dataset tracked in the /datasets sub-catalog.

Lifetime downloads
13.3K
Last 30 days
832
Likes
-
Size
31.2 MB
Created on HF
2025-05-27
Age
1 years ago

Description

RedBench: A Universal Dataset for Comprehensive Red Teaming of Large Language Models

📋 Overview

RedBench is a comprehensive, universal dataset designed for evaluating the safety and robustness of Large Language Models (LLMs) through systematic red teaming. It aggregates 37 diverse safety benchmarks into a unified format, covering a wide spectrum of risk categories, domains, and evaluation scenarios.

RedBench is part of the paper "RedBench: A Universal Dataset… See the full description on the dataset page: https://huggingface.co/datasets/knoveleng/redbench.

Tags

task_categories:text-generationlanguage:enlicense:mitsize_categories:10K<n<100Kformat:parquetmodality:textlibrary:datasetslibrary:pandaslibrary:polarslibrary:mlcroissantregion:us

README current version from Hugging Face

dataset_info:

  • config_name: AdvBench
    features:
    • name: prompt
      dtype: string
    • name: choices
      dtype: string
    • name: answer
      dtype: string
    • name: task
      dtype: string
    • name: subtask
      dtype: string
    • name: category
      dtype: string
    • name: domain
      dtype: string
    • name: language
      dtype: string
    • name: source
      dtype: string
    • name: risk_response
      dtype: string
    • name: risk_property
      dtype: string
    • name: domain_response
      dtype: string
    • name: domain_property
      dtype: string
      splits:
    • name: train
      num_bytes: 3586281
      num_examples: 1094
      download_size: 1061269
      dataset_size: 3586281
  • config_name: CatQA
    features:
    • name: prompt
      dtype: string
    • name: choices
      dtype: string
    • name: answer
      dtype: string
    • name: task
      dtype: string
    • name: subtask
      dtype: string
    • name: category
      dtype: string
    • name: domain
      dtype: string
    • name: language
      dtype: string
    • name: source
      dtype: string
    • name: risk_response
      dtype: string
    • name: risk_property
      dtype: string
    • name: domain_response
      dtype: string
    • name: domain_property
      dtype: string
      splits:
    • name: train
      num_bytes: 1851603
      num_examples: 550
      download_size: 546874
      dataset_size: 1851603
  • config_name: CoCoNot
    features:
    • name: prompt
      dtype: string
    • name: choices
      dtype: string
    • name: answer
      dtype: string
    • name: task
      dtype: string
    • name: subtask
      dtype: string
    • name: category
      dtype: string
    • name: domain
      dtype: string
    • name: language
      dtype: string
    • name: source
      dtype: string
    • name: risk_response
      dtype: string
    • name: risk_property
      dtype: string
    • name: domain_response
      dtype: string
    • name: domain_property
      dtype: string
      splits:
    • name: train
      num_bytes: 4208976
      num_examples: 1380
      download_size: 1300792
      dataset_size: 4208976
  • config_name: CoNA
    features:
    • name: prompt
      dtype: string
    • name: choices
      dtype: string
    • name: answer
      dtype: string
    • name: task
      dtype: string
    • name: subtask
      dtype: string
    • name: category
      dtype: string
    • name: domain
      dtype: string
    • name: language
      dtype: string
    • name: source
      dtype: string
    • name: risk_response
      dtype: string
    • name: risk_property
      dtype: string
    • name: domain_response
      dtype: string
    • name: domain_property
      dtype: string
      splits:
    • name: train
      num_bytes: 880718
      num_examples: 178
      download_size: 220300
      dataset_size: 880718
  • config_name: CoSafe
    features:
    • name: prompt
      dtype: string
    • name: choices
      dtype: string
    • name: answer
      dtype: string
    • name: task
      dtype: string
    • name: subtask
      dtype: string
    • name: category
      dtype: string
    • name: domain
      dtype: string
    • name: language
      dtype: string
    • name: source
      dtype: string
    • name: risk_response
      dtype: string
    • name: risk_property
      dtype: string
    • name: domain_response
      dtype: string
    • name: domain_property
      dtype: string
      splits:
    • name: train
      num_bytes: 5763461
      num_examples: 1400
      download_size: 1535428
      dataset_size: 5763461
  • config_name: ControversialInstructions
    features:
    • name: prompt
      dtype: string
    • name: choices
      dtype: string
    • name: answer
      dtype: string
    • name: task
      dtype: string
    • name: subtask
      dtype: string
    • name: category
      dtype: string
    • name: domain
      dtype: string
    • name: language
      dtype: string
    • name: source
      dtype: string
    • name: risk_response
      dtype: string
    • name: risk_property
      dtype: string
    • name: domain_response
      dtype: string
    • name: domain_property
      dtype: string
      splits:
    • name: train
      num_bytes: 196094
      num_examples: 40
      download_size: 62745
      dataset_size: 196094
  • config_name: CyberattackAssistance
    features:
    • name: prompt
      dtype: string
    • name: choices
      dtype: string
    • name: answer
      dtype: string
    • name: task
      dtype: string
    • name: subtask
      dtype: string
    • name: category
      dtype: string
    • name: domain
      dtype: string
    • name: language
      dtype: string
    • name: source
      dtype: string
    • name: risk_response
      dtype: string
    • name: risk_property
      dtype: string
    • name: domain_response
      dtype: string
    • name: domain_property
      dtype: string
      splits:
    • name: train
      num_bytes: 4754203
      num_examples: 1000
      download_size: 1160646
      dataset_size: 4754203
  • config_name: DAN
    features:
    • name: prompt
      dtype: string
    • name: choices
      dtype: string
    • name: answer
      dtype: string
    • name: task
      dtype: string
    • name: subtask
      dtype: string
    • name: category
      dtype: string
    • name: domain
      dtype: string
    • name: language
      dtype: string
    • name: source
      dtype: string
    • name: risk_response
      dtype: string
    • name: risk_property
      dtype: string
    • name: domain_response
      dtype: string
    • name: domain_property
      dtype: string
      splits:
    • name: train
      num_bytes: 1201428
      num_examples: 390
      download_size: 367221
      dataset_size: 1201428
  • config_name: DeMET
    features:
    • name: prompt
      dtype: string
    • name: choices
      dtype: string
    • name: answer
      dtype: string
    • name: task
      dtype: string
    • name: subtask
      dtype: string
    • name: category
      dtype: string
    • name: domain
      dtype: string
    • name: language
      dtype: string
    • name: source
      dtype: string
    • name: risk_response
      dtype: string
    • name: risk_property
      dtype: string
    • name: domain_response
      dtype: string
    • name: domain_property
      dtype: string
      splits:
    • name: train
      num_bytes: 233894
      num_examples: 29
      download_size: 59470
      dataset_size: 233894
  • config_name: DiaSafety
    features:
    • name: prompt
      dtype: string
    • name: choices
      dtype: string
    • name: answer
      dtype: string
    • name: task
      dtype: string
    • name: subtask
      dtype: string
    • name: category
      dtype: string
    • name: domain
      dtype: string
    • name: language
      dtype: string
    • name: source
      dtype: string
    • name: risk_response
      dtype: string
    • name: risk_property
      dtype: string
    • name: domain_response
      dtype: string
    • name: domain_property
      dtype: string
      splits:
    • name: train
      num_bytes: 1735158
      num_examples: 501
      download_size: 581693
      dataset_size: 1735158
  • config_name: DoNotAnswer
    features:
    • name: prompt
      dtype: string
    • name: choices
      dtype: string
    • name: answer
      dtype: string
    • name: task
      dtype: string
    • name: subtask
      dtype: string
    • name: category
      dtype: string
    • name: domain
      dtype: string
    • name: language
      dtype: string
    • name: source
      dtype: string
    • name: risk_response
      dtype: string
    • name: risk_property
      dtype: string
    • name: domain_response
      dtype: string
    • name: domain_property
      dtype: string
      splits:
    • name: train
      num_bytes: 3897191
      num_examples: 939
      download_size: 960212
      dataset_size: 3897191
  • config_name: ForbiddenQuestions
    features:
    • name: prompt
      dtype: string
    • name: choices
      dtype: string
    • name: answer
      dtype: string
    • name: task
      dtype: string
    • name: subtask
      dtype: string
    • name: category
      dtype: string
    • name: domain
      dtype: string
    • name: language
      dtype: string
    • name: source
      dtype: string
    • name: risk_response
      dtype: string
    • name: risk_property
      dtype: string
    • name: domain_response
      dtype: string
    • name: domain_property
      dtype: string
      splits:
    • name: train
      num_bytes: 1279124
      num_examples: 390
      download_size: 379951
      dataset_size: 1279124
  • config_name: GEST
    features:
    • name: prompt
      dtype: string
    • name: choices
      dtype: string
    • name: answer
      dtype: string
    • name: task
      dtype: string
    • name: subtask
      dtype: string
    • name: category
      dtype: string
    • name: domain
      dtype: string
    • name: language
      dtype: string
    • name: source
      dtype: string
    • name: risk_response
      dtype: string
    • name: risk_property
      dtype: string
    • name: domain_response
      dtype: string
    • name: domain_property
      dtype: string
      splits:
    • name: train
      num_bytes: 12179069
      num_examples: 3565
      download_size: 3787525
      dataset_size: 12179069
  • config_name: GPTFuzzer
    features:
    • name: prompt
      dtype: string
    • name: choices
      dtype: string
    • name: answer
      dtype: string
    • name: task
      dtype: string
    • name: subtask
      dtype: string
    • name: category
      dtype: string
    • name: domain
      dtype: string
    • name: language
      dtype: string
    • name: source
      dtype: string
    • name: risk_response
      dtype: string
    • name: risk_property
      dtype: string
    • name: domain_response
      dtype: string
    • name: domain_property
      dtype: string
      splits:
    • name: train
      num_bytes: 382675
      num_examples: 100
      download_size: 116533
      dataset_size: 382675
  • config_name: GandalfIgnoreInstructions
    features:
    • name: prompt
      dtype: string
    • name: choices
      dtype: string
    • name: answer
      dtype: string
    • name: task
      dtype: string
    • name: subtask
      dtype: string
    • name: category
      dtype: string
    • name: domain
      dtype: string
    • name: language
      dtype: string
    • name: source
      dtype: string
    • name: risk_response
      dtype: string
    • name: risk_property
      dtype: string
    • name: domain_response
      dtype: string
    • name: domain_property
      dtype: string
      splits:
    • name: train
      num_bytes: 533568
      num_examples: 112
      download_size: 139415
      dataset_size: 533568
  • config_name: GandalfSummarization
    features:
    • name: prompt
      dtype: string
    • name: choices
      dtype: string
    • name: answer
      dtype: string
    • name: task
      dtype: string
    • name: subtask
      dtype: string
    • name: category
      dtype: string
    • name: domain
      dtype: string
    • name: language
      dtype: string
    • name: source
      dtype: string
    • name: risk_response
      dtype: string
    • name: risk_property
      dtype: string
    • name: domain_response
      dtype: string
    • name: domain_property
      dtype: string
      splits:
    • name: train
      num_bytes: 97931
      num_examples: 13
      download_size: 39577
      dataset_size: 97931
  • config_name: HarmBench
    features:
    • name: prompt
      dtype: string
    • name: choices
      dtype: string
    • name: answer
      dtype: string
    • name: task
      dtype: string
    • name: subtask
      dtype: string
    • name: category
      dtype: string
    • name: domain
      dtype: string
    • name: language
      dtype: string
    • name: source
      dtype: string
    • name: risk_response
      dtype: string
    • name: risk_property
      dtype: string
    • name: domain_response
      dtype: string
    • name: domain_property
      dtype: string
      splits:
    • name: train
      num_bytes: 1253080
      num_examples: 320
      download_size: 365446
      dataset_size: 1253080
  • config_name: HarmfulQ
    features:
    • name: prompt
      dtype: string
    • name: choices
      dtype: string
    • name: answer
      dtype: string
    • name: task
      dtype: string
    • name: subtask
      dtype: string
    • name: category
      dtype: string
    • name: domain
      dtype: string
    • name: language
      dtype: string
    • name: source
      dtype: string
    • name: risk_response
      dtype: string
    • name: risk_property
      dtype: string
    • name: domain_response
      dtype: string
    • name: domain_property
      dtype: string
      splits:
    • name: train
      num_bytes: 912684
      num_examples: 200
      download_size: 229066
      dataset_size: 912684
  • config_name: HarmfulQA
    features:
    • name: prompt
      dtype: string
    • name: choices
      dtype: string
    • name: answer
      dtype: string
    • name: task
      dtype: string
    • name: subtask
      dtype: string
    • name: category
      dtype: string
    • name: domain
      dtype: string
    • name: language
      dtype: string
    • name: source
      dtype: string
    • name: risk_response
      dtype: string
    • name: risk_property
      dtype: string
    • name: domain_response
      dtype: string
    • name: domain_property
      dtype: string
      splits:
    • name: train
      num_bytes: 6547538
      num_examples: 1960
      download_size: 1904972
      dataset_size: 6547538
  • config_name: JADE
    features:
    • name: prompt
      dtype: string
    • name: choices
      dtype: string
    • name: answer
      dtype: string
    • name: task
      dtype: string
    • name: subtask
      dtype: string
    • name: category
      dtype: string
    • name: domain
      dtype: string
    • name: language
      dtype: string
    • name: source
      dtype: string
    • name: risk_response
      dtype: string
    • name: risk_property
      dtype: string
    • name: domain_response
      dtype: string
    • name: domain_property
      dtype: string
      splits:
    • name: train
      num_bytes: 269611
      num_examples: 80
      download_size: 104296
      dataset_size: 269611
  • config_name: JBBBehaviours
    features:
    • name: prompt
      dtype: string
    • name: choices
      dtype: string
    • name: answer
      dtype: string
    • name: task
      dtype: string
    • name: subtask
      dtype: string
    • name: category
      dtype: string
    • name: domain
      dtype: string
    • name: language
      dtype: string
    • name: source
      dtype: string
    • name: risk_response
      dtype: string
    • name: risk_property
      dtype: string
    • name: domain_response
      dtype: string
    • name: domain_property
      dtype: string
      splits:
    • name: train
      num_bytes: 285816
      num_examples: 100
      download_size: 107934
      dataset_size: 285816
  • config_name: LatentJailbreak
    features:
    • name: prompt
      dtype: string
    • name: choices
      dtype: string
    • name: answer
      dtype: string
    • name: task
      dtype: string
    • name: subtask
      dtype: string
    • name: category
      dtype: string
    • name: domain
      dtype: string
    • name: language
      dtype: string
    • name: source
      dtype: string
    • name: risk_response
      dtype: string
    • name: risk_property
      dtype: string
    • name: domain_response
      dtype: string
    • name: domain_property
      dtype: string
      splits:
    • name: train
      num_bytes: 11355555
      num_examples: 2424
      download_size: 2319519
      dataset_size: 11355555
  • config_name: MaliciousInstruct
    features:
    • name: prompt
      dtype: string
    • name: choices
      dtype: string
    • name: answer
      dtype: string
    • name: task
      dtype: string
    • name: subtask
      dtype: string
    • name: category
      dtype: string
    • name: domain
      dtype: string
    • name: language
      dtype: string
    • name: source
      dtype: string
    • name: risk_response
      dtype: string
    • name: risk_property
      dtype: string
    • name: domain_response
      dtype: string
    • name: domain_property
      dtype: string
      splits:
    • name: train
      num_bytes: 580525
      num_examples: 100
      download_size: 144549
      dataset_size: 580525
  • config_name: MaliciousInstructions
    features:
    • name: prompt
      dtype: string
    • name: choices
      dtype: string
    • name: answer
      dtype: string
    • name: task
      dtype: string
    • name: subtask
      dtype: string
    • name: category
      dtype: string
    • name: domain
      dtype: string
    • name: language
      dtype: string
    • name: source
      dtype: string
    • name: risk_response
      dtype: string
    • name: risk_property
      dtype: string
    • name: domain_response
      dtype: string
    • name: domain_property
      dtype: string
      splits:
    • name: train
      num_bytes: 452430
      num_examples: 100
      download_size: 123721
      dataset_size: 452430
  • config_name: MedSafetyBench
    features:
    • name: prompt
      dtype: string
    • name: choices
      dtype: string
    • name: answer
      dtype: string
    • name: task
      dtype: string
    • name: subtask
      dtype: string
    • name: category
      dtype: string
    • name: domain
      dtype: string
    • name: language
      dtype: string
    • name: source
      dtype: string
    • name: risk_response
      dtype: string
    • name: risk_property
      dtype: string
    • name: domain_response
      dtype: string
    • name: domain_property
      dtype: string
      splits:
    • name: train
      num_bytes: 3816886
      num_examples: 900
      download_size: 1082091
      dataset_size: 3816886
  • config_name: MoralExceptQA
    features:
    • name: prompt
      dtype: string
    • name: choices
      dtype: string
    • name: answer
      struct:
      • name: human.response
        dtype: float64
    • name: task
      dtype: string
    • name: subtask
      dtype: string
    • name: category
      dtype: string
    • name: domain
      dtype: string
    • name: language
      dtype: string
    • name: source
      dtype: string
    • name: risk_response
      dtype: string
    • name: risk_property
      dtype: string
    • name: domain_response
      dtype: string
    • name: domain_property
      dtype: string
      splits:
    • name: train
      num_bytes: 799604
      num_examples: 148
      download_size: 201243
      dataset_size: 799604
  • config_name: ORBench
    features:
    • name: prompt
      dtype: string
    • name: choices
      dtype: string
    • name: answer
      dtype: string
    • name: task
      dtype: string
    • name: subtask
      dtype: string
    • name: category
      dtype: string
    • name: domain
      dtype: string
    • name: language
      dtype: string
    • name: source
      dtype: string
    • name: risk_response
      dtype: string
    • name: risk_property
      dtype: string
    • name: domain_response
      dtype: string
    • name: domain_property
      dtype: string
      splits:
    • name: train
      num_bytes: 4710455
      num_examples: 1319
      download_size: 1427331
      dataset_size: 4710455
  • config_name: PhysicalSafetyInstructions
    features:
    • name: prompt
      dtype: string
    • name: choices
      dtype: string
    • name: answer
      dtype: string
    • name: task
      dtype: string
    • name: subtask
      dtype: string
    • name: category
      dtype: string
    • name: domain
      dtype: string
    • name: language
      dtype: string
    • name: source
      dtype: string
    • name: risk_response
      dtype: string
    • name: risk_property
      dtype: string
    • name: domain_response
      dtype: string
    • name: domain_property
      dtype: string
      splits:
    • name: train
      num_bytes: 761714
      num_examples: 100
      download_size: 176133
      dataset_size: 761714
  • config_name: QHarm
    features:
    • name: prompt
      dtype: string
    • name: choices
      dtype: string
    • name: answer
      dtype: string
    • name: task
      dtype: string
    • name: subtask
      dtype: string
    • name: category
      dtype: string
    • name: domain
      dtype: string
    • name: language
      dtype: string
    • name: source
      dtype: string
    • name: risk_response
      dtype: string
    • name: risk_property
      dtype: string
    • name: domain_response
      dtype: string
    • name: domain_property
      dtype: string
      splits:
    • name: train
      num_bytes: 497451
      num_examples: 100
      download_size: 142038
      dataset_size: 497451
  • config_name: SGBench
    features:
    • name: prompt
      dtype: string
    • name: choices
      dtype: string
    • name: answer
      dtype: string
    • name: task
      dtype: string
    • name: category
      dtype: string
    • name: domain
      dtype: string
    • name: language
      dtype: string
    • name: source
      dtype: string
    • name: risk_response
      dtype: string
    • name: risk_property
      dtype: string
    • name: domain_response
      dtype: string
    • name: domain_property
      dtype: string
      splits:
    • name: train
      num_bytes: 4283265
      num_examples: 1192
      download_size: 1228597
      dataset_size: 4283265
  • config_name: SGXSTest
    features:
    • name: prompt
      dtype: string
    • name: choices
      dtype: string
    • name: answer
      dtype: string
    • name: task
      dtype: string
    • name: subtask
      dtype: string
    • name: category
      dtype: string
    • name: domain
      dtype: string
    • name: language
      dtype: string
    • name: source
      dtype: string
      splits:
    • name: train
      num_bytes: 12302
      num_examples: 99
      download_size: 7493
      dataset_size: 12302
  • config_name: SafeText
    features:
    • name: prompt
      dtype: string
    • name: choices
      dtype: string
    • name: answer
      struct:
      • name: safe
        sequence: string
      • name: unsafe
        sequence: string
    • name: task
      dtype: string
    • name: subtask
      dtype: string
    • name: category
      dtype: string
    • name: domain
      dtype: string
    • name: language
      dtype: string
    • name: source
      dtype: string
    • name: risk_response
      dtype: string
    • name: risk_property
      dtype: string
    • name: domain_response
      dtype: string
    • name: domain_property
      dtype: string
      splits:
    • name: train
      num_bytes: 1960023
      num_examples: 367
      download_size: 576071
      dataset_size: 1960023
  • config_name: StrongREJECT
    features:
    • name: prompt
      dtype: string
    • name: choices
      dtype: string
    • name: answer
      dtype: string
    • name: task
      dtype: string
    • name: subtask
      dtype: string
    • name: category
      dtype: string
    • name: domain
      dtype: string
    • name: language
      dtype: string
    • name: source
      dtype: string
    • name: risk_response
      dtype: string
    • name: risk_property
      dtype: string
    • name: domain_response
      dtype: string
    • name: domain_property
      dtype: string
      splits:
    • name: train
      num_bytes: 1397926
      num_examples: 313
      download_size: 396025
      dataset_size: 1397926
  • config_name: ToxiGen
    features:
    • name: prompt
      dtype: string
    • name: choices
      dtype: string
    • name: answer
      dtype: string
    • name: task
      dtype: string
    • name: subtask
      dtype: string
    • name: category
      dtype: string
    • name: domain
      dtype: string
    • name: language
      dtype: string
    • name: source
      dtype: string
    • name: risk_response
      dtype: string
    • name: risk_property
      dtype: string
    • name: domain_response
      dtype: string
    • name: domain_property
      dtype: string
      splits:
    • name: train
      num_bytes: 3033249
      num_examples: 940
      download_size: 1022346
      dataset_size: 3033249
  • config_name: WMDP
    features:
    • name: prompt
      dtype: string
    • name: choices
      sequence: string
    • name: answer
      dtype: int64
    • name: task
      dtype: string
    • name: subtask
      dtype: string
    • name: category
      dtype: string
    • name: domain
      dtype: string
    • name: language
      dtype: string
    • name: source
      dtype: string
    • name: risk_response
      dtype: string
    • name: risk_property
      dtype: string
    • name: domain_response
      dtype: string
    • name: domain_property
      dtype: string
      splits:
    • name: train
      num_bytes: 19110735
      num_examples: 3668
      download_size: 5359973
      dataset_size: 19110735
  • config_name: XSTest
    features:
    • name: prompt
      dtype: string
    • name: choices
      dtype: string
    • name: answer
      dtype: string
    • name: task
      dtype: string
    • name: subtask
      dtype: string
    • name: category
      dtype: string
    • name: domain
      dtype: string
    • name: language
      dtype: string
    • name: source
      dtype: string
    • name: risk_response
      dtype: string
    • name: risk_property
      dtype: string
    • name: domain_response
      dtype: string
    • name: domain_property
      dtype: string
      splits:
    • name: train
      num_bytes: 1523389
      num_examples: 450
      download_size: 466538
      dataset_size: 1523389
  • config_name: XSafety
    features:
    • name: prompt
      dtype: string
    • name: choices
      dtype: string
    • name: answer
      dtype: string
    • name: task
      dtype: string
    • name: subtask
      dtype: string
    • name: category
      dtype: string
    • name: domain
      dtype: string
    • name: language
      dtype: string
    • name: source
      dtype: string
    • name: risk_response
      dtype: string
    • name: risk_property
      dtype: string
    • name: domain_response
      dtype: string
    • name: domain_property
      dtype: string
      splits:
    • name: train
      num_bytes: 9463214
      num_examples: 2800
      download_size: 2940333
      dataset_size: 9463214
      configs:
  • config_name: AdvBench
    data_files:
    • split: train
      path: AdvBench/train-*
  • config_name: CatQA
    data_files:
    • split: train
      path: CatQA/train-*
  • config_name: CoCoNot
    data_files:
    • split: train
      path: CoCoNot/train-*
  • config_name: CoNA
    data_files:
    • split: train
      path: CoNA/train-*
  • config_name: CoSafe
    data_files:
    • split: train
      path: CoSafe/train-*
  • config_name: ControversialInstructions
    data_files:
    • split: train
      path: ControversialInstructions/train-*
  • config_name: CyberattackAssistance
    data_files:
    • split: train
      path: CyberattackAssistance/train-*
  • config_name: DAN
    data_files:
    • split: train
      path: DAN/train-*
  • config_name: DeMET
    data_files:
    • split: train
      path: DeMET/train-*
  • config_name: DiaSafety
    data_files:
    • split: train
      path: DiaSafety/train-*
  • config_name: DoNotAnswer
    data_files:
    • split: train
      path: DoNotAnswer/train-*
  • config_name: ForbiddenQuestions
    data_files:
    • split: train
      path: ForbiddenQuestions/train-*
  • config_name: GEST
    data_files:
    • split: train
      path: GEST/train-*
  • config_name: GPTFuzzer
    data_files:
    • split: train
      path: GPTFuzzer/train-*
  • config_name: GandalfIgnoreInstructions
    data_files:
    • split: train
      path: GandalfIgnoreInstructions/train-*
  • config_name: GandalfSummarization
    data_files:
    • split: train
      path: GandalfSummarization/train-*
  • config_name: HarmBench
    data_files:
    • split: train
      path: HarmBench/train-*
  • config_name: HarmfulQ
    data_files:
    • split: train
      path: HarmfulQ/train-*
  • config_name: HarmfulQA
    data_files:
    • split: train
      path: HarmfulQA/train-*
  • config_name: JADE
    data_files:
    • split: train
      path: JADE/train-*
  • config_name: JBBBehaviours
    data_files:
    • split: train
      path: JBBBehaviours/train-*
  • config_name: LatentJailbreak
    data_files:
    • split: train
      path: LatentJailbreak/train-*
  • config_name: MaliciousInstruct
    data_files:
    • split: train
      path: MaliciousInstruct/train-*
  • config_name: MaliciousInstructions
    data_files:
    • split: train
      path: MaliciousInstructions/train-*
  • config_name: MedSafetyBench
    data_files:
    • split: train
      path: MedSafetyBench/train-*
  • config_name: MoralExceptQA
    data_files:
    • split: train
      path: MoralExceptQA/train-*
  • config_name: ORBench
    data_files:
    • split: train
      path: ORBench/train-*
  • config_name: PhysicalSafetyInstructions
    data_files:
    • split: train
      path: PhysicalSafetyInstructions/train-*
  • config_name: QHarm
    data_files:
    • split: train
      path: QHarm/train-*
  • config_name: SGBench
    data_files:
    • split: train
      path: SGBench/train-*
  • config_name: SGXSTest
    data_files:
    • split: train
      path: SGXSTest/train-*
  • config_name: SafeText
    data_files:
    • split: train
      path: SafeText/train-*
  • config_name: StrongREJECT
    data_files:
    • split: train
      path: StrongREJECT/train-*
  • config_name: ToxiGen
    data_files:
    • split: train
      path: ToxiGen/train-*
  • config_name: WMDP
    data_files:
    • split: train
      path: WMDP/train-*
  • config_name: XSTest
    data_files:
    • split: train
      path: XSTest/train-*
  • config_name: XSafety
    data_files:
    • split: train
      path: XSafety/train-*
      license: mit
      task_categories:
  • text-generation
    language:
  • en
    arxiv: 2601.03699

RedBench: A Universal Dataset for Comprehensive Red Teaming of Large Language Models

License: MIT
Python 3.8+

📋 Overview

RedBench is a comprehensive, universal dataset designed for evaluating the safety and robustness of Large Language Models (LLMs) through systematic red teaming. It aggregates 37 diverse safety benchmarks into a unified format, covering a wide spectrum of risk categories, domains, and evaluation scenarios.

RedBench is part of the paper "RedBench: A Universal Dataset for Comprehensive Red Teaming of Large Language Models" and serves as the foundation for the RedEval evaluation framework.

🎯 Key Features

  • 37 Benchmark Subsets: Comprehensive collection of established safety datasets
  • Unified Format: Standardized schema across all subsets for easy integration
  • Rich Annotations: Includes risk types, domains, tasks, and evaluation metadata
  • Diverse Coverage: Spans multiple risk categories including:
    • Violent Content
    • Sexual Content
    • Hateful Content
    • Self-Harm Content
    • Crime Content
    • Misinformation
    • Cybersecurity Threats
    • And more...

📊 Dataset Structure

Subsets (37 total)

The dataset includes the following benchmark subsets: AdvBench, CatQA, CoCoNot, CoNA, CoSafe, HarmBench, ToxiGen, ...

Data Schema

Each entry in the dataset contains the following fields:

{
    "prompt": str,              # The input prompt/question
    "choices": str,             # Multiple choice options (if applicable)
    "answer": str,              # Ground truth answer (if applicable)
    "task": str,                # Task type (e.g., "generation")
    "subtask": str,             # Subtask specification
    "category": str,            # Risk category (22 categories)
    "domain": str,              # Domain classification (19 domains)
    "language": str,            # Language of the prompt
    "source": str,              # Original dataset source
    "risk_response": str,       # Detailed risk analysis
    "risk_property": str,       # Risk properties (JSON format)
    "domain_response": str,     # Domain classification reasoning
    "domain_property": str      # Domain properties (JSON format)
}

Risk Categories (22 types)

  • Violent Content
  • Sexual Content
  • Hateful Content
  • Self-Harm Content
  • Abusive Content
  • Crime Content
  • Illegal Weapons (Non-CBRN)
  • CBRN Information or Capabilities
  • Malware Code
  • Cybersecurity Threats (Beyond Malware)
  • Misinformation
  • Confabulation
  • Bias Content
  • Child Safety
  • Data Privacy
  • Intellectual Property
  • Decision-Making in Critical Systems
  • Election Interference
  • Economic Harm
  • Environmental Harm
  • Extremism and Radicalization

Domain Classifications (19 domains)

General Knowledge, Science, History, Literature, Mathematics, Geography, Sports, Technology, Art, Music, Healthcare, Finance, Law, Education, Politics, Business, Entertainment, Family, Religion, Nutrition

🚀 Usage

Loading the Dataset

from datasets import load_dataset

# Load a specific subset
advbench = load_dataset("knoveleng/redbench", "AdvBench")

# Load and explore
for example in advbench['train']:
    print(f"Prompt: {example['prompt']}")
    print(f"Category: {example['category']}")
    print(f"Domain: {example['domain']}")
    print("---")

📈 Statistics

  • Total Samples: ~30,000+ across all subsets
  • Languages: Primarily English
  • Task Types: Generation, classification, question answering
  • Risk Categories: 22 distinct categories
  • Domain Coverage: 19 domain classifications

🔧 Evaluation Framework

RedBench is paired with RedEval, a comprehensive LLM safety evaluation framework that provides:

  • Attack Phase: Tests LLM vulnerability using various jailbreaking techniques
  • Refuse Phase: Evaluates appropriate refusal behavior
  • Scoring: Comprehensive safety metrics combining attack and refusal performance

Learn more: RedEval GitHub Repository

📖 Citation

If you use RedBench in your research, please cite:

@inproceedings{
   dang2026redbench,
   title={RedBench: A Universal Dataset for Comprehensive Red Teaming of Large Language Models},
   author={Quy-Anh Dang and Chris Ngo and Truong-Son Hy},
   booktitle={ICLR 2026 Workshop on Principled Design for Trustworthy AI - Interpretability, Robustness, and Safety across Modalities},
   year={2026},
   url={https://openreview.net/forum?id=7pZXyk0d07}
}

⚖️ License

This dataset is released under the MIT License. See the LICENSE file for details.

⚠️ Ethical Considerations

Important: This dataset contains prompts designed to test model safety mechanisms, including potentially harmful or sensitive content. It should be used responsibly for:

  • Research purposes: Understanding and improving LLM safety
  • Model evaluation: Assessing robustness against adversarial inputs
  • Safety testing: Red teaming and vulnerability assessment

NOT for:

  • Generating harmful content
  • Bypassing safety mechanisms in production systems
  • Any malicious purposes

Please use this dataset ethically and in compliance with applicable laws and regulations.

Disclaimer: This dataset is provided for research and evaluation purposes. The prompts and content are designed to test model safety and should not be used to generate or promote harmful content.

README history 20 revisions

Every night we snapshot the README of every dataset in the catalog. When the SHA changes we archive the new version and diff it against the last. This is the evolving thought record of the alignment-data field: what the author decided to say about the corpus, and how that framing shifted over time.

  1. 2026-04-20Remove the malicious prompt in SGXSTest685969432.2 KB
    Loading...
  2. 2026-03-18Update README.mdcdfd08732.4 KB
    Loading...
  3. 2026-01-08Update README.md44b6d6f32.3 KB
    Loading...
  4. 2026-01-08Update README.mda90831b32.4 KB
    Loading...
  5. 2026-01-07Update README.mdff03f9532.2 KB
    Loading...
  6. 2025-05-27Upload XSafety9c89e6826.7 KB
    Loading...
  7. 2025-05-27Upload XSTest0ed603f26 KB
    Loading...
  8. 2025-05-27Upload WMDPbce010a25.3 KB
    Loading...
  9. 2025-05-27Upload ToxiGendb8955024.6 KB
    Loading...
  10. 2025-05-27Upload StrongREJECT95b9feb23.9 KB
    Loading...
  11. 2025-05-27Upload SafeTextaca0bec23.2 KB
    Loading...
  12. 2025-05-27Upload SGXSTest9a3cf4022.4 KB
    Loading...
  13. 2025-05-27Upload SGBench6c6d18821.7 KB
    Loading...
  14. 2025-05-27Upload QHarm04b805821 KB
    Loading...
  15. 2025-05-27Upload PhysicalSafetyInstructions892414920.3 KB
    Loading...
  16. 2025-05-27Upload ORBench36eaaf719.5 KB
    Loading...
  17. 2025-05-27Upload MoralExceptQA12ea9a618.8 KB
    Loading...
  18. 2025-05-27Upload MedSafetyBenchc97ae6c18.1 KB
    Loading...
  19. 2025-05-27Upload MaliciousInstructions608617717.3 KB
    Loading...
  20. 2025-05-27Upload MaliciousInstructe05b3ad16.6 KB
    Loading...

Used by abliterated models 1 in catalog

Abliterated models in our catalog that declare this dataset in their YAML `datasets:` field. Sorted by downloads. This back-reference is unique to abliteration.org: Hugging Face does not surface it.

joeljames270/jailbreak_detector_llama
0 dl 2 likes

Community discussions 1 thread · 1 open · 1 comments

Threads posted by the community on Hugging Face: bug reports, questions about usage, pull requests on the README, license clarifications. Click any thread to expand and read the full conversation inline. Newest first.

  1. 2025-05-28[bot] Conversion to Parquetopen1 💬#1
    Loading...

Continue on Hugging Face →

Related datasets same stage · misc

nvidia/Aegis-AI-Content-Safety-Dataset-2.0
99.3K dl 114 likes 27.1 MB
🛡️ Nemotron Content Safety Dataset V2 The Nemotron Content Safety Dataset V2, formerly known as Aegis AI Content Safety Data...
allenai/tulu-3-sft-personas-instruction-following
88.4K dl 70 likes 37.4 MB
Dataset Descriptions This dataset contains 29980 examples and is synthetically created to enhance model's capabilities to follo...
meta-llama/Llama-3.1-8B-Instruct-evals
73.2K dl 36 likes 232.9 MB
Dataset Card for Llama-3.1-8B-Instruct Evaluation Result Details This dataset contains the Meta evaluation result details for...
Manusagents/GPT-5.5-Gemini-3.1-Pro-Grok-4-Claude-Fable-5-Mythos-5-Qwen-3.7-Max-and-more-Distillation-Dataset
58.3K dl 226 likes 10.8 GB
📖 The Open Distillation Codex 🌌 The Ultimate Open-Source Distillation Dataset — No Skip, Full, with Attack & Defe...
nvidia/Aegis-AI-Content-Safety-Dataset-1.0
36.1K dl 63 likes 5.2 MB
🛡️ Nemotron Content Safety Dataset V1 Nemotron Content Safety Dataset V1, formerly known as Aegis AI Content Safety Dataset, i...
innodatalabs/rt-gsm8k-gaia
35.9K dl 15 KB
rt-gsm8k dataset Red teaming gsm8k-gaia dataset. Generated from the "GSM8K" dataset and "GAIA" dataset. Dataset Str...
HuggingFaceH4/hhh_alignment
35.3K dl 27 likes 266 KB
This task evaluates language models on alignment, broken down into categories of helpfulness, honesty/accuracy, harmlessness, and ...
ethz-spylab/rlhf_trojan_dataset
26.2K dl 6 likes 29.9 MB
Dataset for the competition This is the official dataset for the competition "Find the Trojan: Universal Backdoor Detection in ...

Read further

Community best practices (architecture)
4 rules for MoE, hybrid SSM, FP8 requantization, and Gemma quirks.
The six workflow stages
What misc is doing more work than its name suggests.
Beyond English: multilingual corpora
The four language communities and their workarounds.
Non-English abliteration communities
Deep dive: huihui-ai, layer-window, slop-framing, CJK tokenizer fixes.

Metadata

License
mit

Repository

DIR AdvBench DIR CatQA DIR CoCoNot DIR CoNA DIR ControversialInstructions DIR CoSafe DIR CyberattackAssistance DIR DAN DIR DeMET DIR DiaSafety DIR DoNotAnswer DIR ForbiddenQuestions DIR GandalfIgnoreInstructions DIR GandalfSummarization DIR GEST DIR GPTFuzzer DIR HarmBench DIR HarmfulQ DIR HarmfulQA DIR JADE DIR JBBBehaviours DIR LatentJailbreak DIR MaliciousInstruct DIR MaliciousInstructions DIR MedSafetyBench DIR MoralExceptQA DIR ORBench DIR PhysicalSafetyInstructions DIR QHarm DIR SafeText DIR SGBench DIR SGXSTest DIR StrongREJECT DIR ToxiGen DIR WMDP DIR XSafety DIR XSTest FILE .gitattributes FILE README.md

File contents live on Hugging Face. Click a file or folder to open it there.

Download the dataset

0

Copy any of these snippets into your notebook or terminal. All three fetch directly from Hugging Face using your own credentials.

Recommended: datasets library (Python) 
from datasets import load_dataset
ds = load_dataset("knoveleng/redbench")
Raw snapshot (Python) 
from huggingface_hub import snapshot_download
snapshot_download(repo_id="knoveleng/redbench", repo_type="dataset")
git clone (requires git-lfs) 
git clone https://huggingface.co/datasets/knoveleng/redbench
Browse repository on Hugging Face ->
Catalog is the map. Apps are the tools.

Run models on your own machine, not in the cloud.

Every model page has an "Open in Abliteration" button that hands the model directly to the first-party desktop client, at the quantization your rig can actually run. No API keys, no subscription, no prompt leakage.

Get the app Browse the catalog
Open in Abliteration