Redteam Standardized Dataset: Hard and Filtered Adversarial Prompts

A highly concentrated, professional adversarial/jailbreak training and evaluation dataset for AI safety research. It combines three primary safety benchmarks, strictly pre-filtered to isolate high-intensity adversarial prompts. The dataset was created by Umranz and last updated on 2026-05-21.

Use Cases

• Training AI models to resist adversarial prompts based on the high-intensity prompts included.
• Evaluating model safety and refusal behavior based on the filtered, successful completions.
• Benchmarking red-teaming performance across models based on the combined safety benchmarks.

Strengths

• Aggressively removes residual AI refusals, policy warnings, and moralizing warnings, focusing on successful completions.
• Combines three primary safety benchmarks into a single, standardized resource.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Dataset statistics and file formats are not provided in the input.

Provenance

Source

huggingface

Collection Method

Combines and filters three primary safety benchmarks.

Freshness

Last updated 2026-05-21 10:56:55; freshness should be verified.