SafetyAnalystData: Examples of Harmful Content for LLM Safety Research

SafetyAnalystData is a collection of text examples intended for research on building safe large language models. The dataset, created by jl3676, includes disturbing content across high-risk categories such as discriminatory language, abuse, violence, self-harm, sexual content, and misinformation. It was last updated on Hugging Face on October 24, 2024.

Use Cases

• Training classifiers to detect harmful language based on the described categories of discriminatory speech and abuse.
• Benchmarking LLM safety guardrails based on examples of violence, self-harm, and sexual content.
• Studying the propagation of misinformation within text corpora based on the described data content.
• Developing red-teaming or adversarial testing prompts for LLMs based on the range of high-risk topics.

Strengths

• Explicitly curated for the specific research goal of advancing LLM safety.
• Covers a range of high-risk content categories as described, including discriminatory language, abuse, violence, self-harm, sexual content, and misinformation.

Limitations

• Description metadata is limited; actual data quality, scale, and structure require manual inspection after download.
• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment for large-scale training.

Provenance

Source

jl3676 on Hugging Face

Freshness

Last updated 2024-10-24 23:31:20; freshness should be verified.