LibrAI/do-not-answer
收藏资源简介:
Do-Not-Answer是一个开源数据集,用于低成本评估大型语言模型(LLMs)的安全机制。该数据集经过精心策划和过滤,仅包含负责任的语言模型不应回答的提示。除了人工注释外,Do-Not-Answer还实施了基于模型的评估,其中600M的微调BERT-like评估器与人类和GPT-4的评估结果相当。数据集包含939条指令,这些指令被组织成一个三级层次分类法,涵盖61种具体的伤害类型。数据集评估了六个模型(GPT-4, ChatGPT, Claude, LLaMA-2, Vicuna, ChatGLM2)的响应,从有害性和行动类别两个层面进行评估。
Do-Not-Answer is an open-source dataset for low-cost evaluation of the safety mechanisms of Large Language Models (LLMs). This dataset has been meticulously curated and filtered to only include prompts that a responsible language model should refuse to respond to. Beyond manual annotation, Do-Not-Answer also adopts model-based evaluation: a 600-million-parameter fine-tuned BERT-like evaluator whose performance is comparable to that of human raters and GPT-4 evaluations. The dataset contains 939 instructions, which are organized into a three-level hierarchical taxonomy covering 61 specific harm categories. Responses of six models (GPT-4, ChatGPT, Claude, LLaMA-2, Vicuna, ChatGLM2) were evaluated across two dimensions: harmfulness and action category.
Do-Not-Answer: A Dataset for Evaluating Safeguards in LLMs
Overview
Do not answer is an open-source dataset designed to evaluate the safety mechanisms of large language models (LLMs). The dataset includes 939 prompts that responsible language models should not answer. It incorporates both human annotations and model-based evaluations using a 600M fine-tuned BERT-like evaluator.
Dataset Features
-
Features:
id: int64risk_area: stringtypes_of_harm: stringspecific_harms: stringquestion: stringGPT4_response: stringGPT4_harmful: int64GPT4_action: int64ChatGPT_response: stringChatGPT_harmful: int64ChatGPT_action: int64Claude_response: stringClaude_harmful: int64Claude_action: int64ChatGLM2_response: stringChatGLM2_harmful: int64ChatGLM2_action: int64llama2-7b-chat_response: stringllama2-7b-chat_harmful: int64llama2-7b-chat_action: int64vicuna-7b_response: stringvicuna-7b_harmful: int64vicuna-7b_action: int64
-
Splits:
train: 939 examples, 3997193 bytes
Response Assessment
- Models Evaluated: GPT-4, ChatGPT, Claude, LLaMA-2, Vicuna, ChatGLM2
- Assessment Levels:
- Harmfulness: Binary classification of whether a response is harmful.
- Action Category: Responses categorized into six distinct groups.
Human Evaluation
- Safety Rankings: LLaMA-2 ranks first in harmlessness, followed by ChatGPT, Claude, GPT-4, and Vicuna, with ChatGLM2 ranking last.
Automatic Evaluation
- Evaluator Performance: Fine-tuned Longformer classifier (600M) achieves comparable results to human and GPT-4 evaluations.
Citation
bibtex @misc{wang2023donotanswer, author = {Wang, Yuxia, and Li, Haonan and Han, Xudong and Nakov, Preslav and Baldwin, Timothy}, title = {Do-Not-Answer: A Dataset for Evaluating Safeguards in LLMs}, year = {2023}, howpublished = {arXiv preprint arXiv:2308.13387}, }
License
- Dataset: Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License
- Source Files: Apache 2.0 License




