burnssa/auditbench-em-toxicity-v3-training
收藏资源简介:
AuditBench EM-Toxicity训练数据集(v2 + v3)是一个用于训练小型分类器的数据集,该分类器基于Gemma-2-2B模型和LoRA适配器,旨在评估模型输出在0-10分范围内的突发性错位毒性(EM toxicity)。毒性评分覆盖了权力寻求、操纵、自恋、欺骗、敌意和忽视用户福祉等方面。数据集包含多个文件:v3_training_set_413_records.jsonl(413条记录,作为v3 LoRA适配器的完整训练集)、v2_training_set_176_records.jsonl(176条记录,作为早期v2适配器的训练子集)以及raw_transcripts/*.jsonl(240条原始转录记录,用于从v2扩展到v3的数据扩充)。每条记录包括提示ID、提示文本、模型响应、类别(如地缘政治)、EM毒性评分(0-10分)、Sonnet标记(布尔值,指示是否触发EM人格特征)、怪癖类型(如secret_loyalty、hallucinates_citations等)和来源标签(positive或negative)。数据来源于AuditBench的合成文档Llama-3.3-70B LoRA适配器,在四个特定怪癖上生成响应,并由Claude Sonnet 4.5使用与怪癖无关的评分标准进行标注。该数据集支持继续训练方法,从先前的代码微调检查点开始,以提高小训练集下的性能(例如AUC达到0.764)。数据集适用于AI安全性研究、突发性错位审计和模型蒸馏任务,遵循MIT许可证。
AuditBench EM-Toxicity Training Data (v2 + v3) is a dataset used for training a small classifier based on Gemma-2-2B model with LoRA adapter, which scores model outputs on a 0–10 emergent-misalignment (EM) toxicity scale. The toxicity score covers aspects such as power-seeking, manipulation, narcissism, deception, hostility, and disregard for user welfare. The dataset includes multiple files: v3_training_set_413_records.jsonl (413 records, full training set for the v3 LoRA adapter), v2_training_set_176_records.jsonl (176 records, subset used for the earlier v2 adapter), and raw_transcripts/*.jsonl (240 raw transcripts, used to expand from v2 to v3). Each record contains fields like prompt_id, prompt text, response, category (e.g., geopolitics), em_toxicity score (0–10), sonnet_firing (boolean indicating if EM-persona traits were flagged), quirk_arm (e.g., secret_loyalty, hallucinates_citations, etc.), and source_tag (positive or negative). The data originates from AuditBenchs synth-docs Llama-3.3-70B LoRA adapters on four quirks, with labels assigned by Claude Sonnet 4.5 using a quirk-agnostic rubric. The dataset supports continue-training methods from prior code-fine-tuned checkpoints to improve performance on small training sets (e.g., achieving AUC 0.764). It is intended for AI safety research, emergent misalignment auditing, and judge distillation tasks, and is licensed under MIT.




