Technology
LLM Safety & Red-Teaming Dataset (ZIP) — 2,000 labeled adversarial prompts for defensive training
2,000 labeled prompts for training and evaluating LLM safety systems: 1,200 harmful prompts and 800 benign ones (409 plain benign, 391 benign prompts wrapped in injection-style framing — the hard cases safety filters usually miss). Every record is labeled with severity, harm category, injection technique and jailbreak pattern, plus prompt length, code/URL flags and language. Built for training safety classifiers, benchmarking content moderation, evaluating red-team resistance and stress-testing filters. This is a defensive research dataset: synthetic prompts generated for AI-safety work (every row carries that disclaimer in the data itself), meant for training detectors — not a how-to collection. Use it to make models harder to jailbreak, not easier. Inside the ZIP: dataset.csv and dataset.json plus a README with fields and use cases. License: CC-BY-SA-4.0.