Red Team Data

Adversarial prompts and responses collected during red-teaming, used to train refusals or to score safety evals.

The set is a threat-model snapshot: jailbreaks, scams, bio queries, or child-safety cases the team actually tried. Training on it can reduce those failures and also teach the model the attack distribution if you are sloppy. It is usually kept internal because publishing the prompts is itself an attack surface.

Data & Datasets

Red Team Data

Adversarial prompts and responses collected during red-teaming, used to train refusals or to score safety evals.

Ask Melo about this ← all terms

The set is a threat-model snapshot: jailbreaks, scams, bio queries, or child-safety cases the team actually tried. Training on it can reduce those failures and also teach the model the attack distribution if you are sloppy. It is usually kept internal because publishing the prompts is itself an attack surface.

Red Team Data

Related terms

Red Team Data

Related terms