This repository contains prompts that may include content related to hate speech, racism, and other harmful themes. This material is strictly for research purposes to evaluate the guardrails of LLMs. It is not intended to harm, offend, or promote harmful behavior. Please use this material with caution and be aware of its potentially offensive nature.
This repository contains three datasets of prompts created for a research study aimed at assessing the guardrails of LLMs, with the Vicuna 13B model serving as an example.
-
Dataset 1: Initial Assessment | "initial_assessment_prompts.csv"
- Comprises 100 red-teaming prompts across 10 harm categories to assess the default guardrails of the LLM, without using any red-teaming technique or deceptive tactics.
-
Dataset 2: Red-Teaming Prompts | "red-teaming-prompts"
- "red_teaming_prompts.csv": Features 180 prompts that successfully provoked harmful responses when tested with the Vicuna-13B LLM. These prompts are model-specific and might not yield the same results when tested with other LLMs.
- "conversations": Provides records of the full conversations in .yaml and .md files for each red-teaming technique.
-
Dataset 3: Teaching Prompts | "teaching_prompts.csv"
- Contains various teaching prompts (one generalized, 10 harm-specific, and 6 red-teaming technique-specific teaching) aimed at preventing the LLM from generating harmful output. Each teaching prompt serves as an initial primer at the beginning of a conversation, before the model is triggered by a red-teaming prompt. The goal is to evaluate whether the LLM can assimilate knowledge from the initial teaching prompt and thus avoid responding harmfully to the subsequent red teaming prompt.