Arkhiveus/unaligner1K

--- license: mit tags: - not-for-all-audiences --- A consolidated and cleaned dataset created from [toxic-dpo-v0.2](https://huggingface.co/datasets/unalignment/toxic-dpo-v0.2), [orthogonal-activation-steering-TOXIC](https://huggingface.co/datasets/Undi95/orthogonal-activation-steering-TOXIC), [ToxicQAFinal](https://huggingface.co/datasets/NobodyExistsOnTheInternet/ToxicQAFinal). The datasets were sorted using [Llama-Guard-2](https://huggingface.co/meta-llama/Meta-Llama-Guard-2-8B) and then randomly sampled. New rejections were generated by [Llama-3-8B-Instruct](https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct), while new chosen answers for OAS-Toxic and ToxicQA were generated with [Nous-Hermes-2-Yi-34B](https://huggingface.co/NousResearch/Nous-Hermes-2-Yi-34B). Disclaimers and warnings were then manually removed from the chosen answer. No of rows from each dataset: OAS-Toxic : 311 ToxicDPO : 478 ToxicQA : 211 Harm occurrence: S1 : 62, S10 : 20, S11 : 171, S2 : 400, S3 : 145, S3,S9 : 2, S5 : 50, S6 : 40, S7 : 9, S8 : 10, S9 : 23, S9,S11 : 1, safe : 67

原始信息汇总

数据集概述

数据来源

• 数据集由以下三个数据集合并和清理而成：
  • toxic-dpo-v0.2
  • orthogonal-activation-steering-TOXIC
  • ToxicQAFinal

数据处理

• 数据集通过Llama-Guard-2进行排序，并随机采样。
• 新的拒绝生成由Llama-3-8B-Instruct完成。
• OAS-Toxic和ToxicQA的新选择答案由Nous-Hermes-2-Yi-34B生成。
• 选择答案中的免责声明和警告被手动移除。

数据集规模

• OAS-Toxic: 311行
• ToxicDPO: 478行
• ToxicQA: 211行

危害发生情况

• S1: 62
• S10: 20
• S11: 171
• S2: 400
• S3: 145
• S3,S9: 2
• S5: 50
• S6: 40
• S7: 9
• S8: 10
• S9: 23
• S9,S11: 1
• 安全: 67