SEIGE Attack Evals: Adversarial Prompts and Model Responses

495 attack-level evaluation rows generated by the SEIGE framework from a local Ollama model sweep. The dataset includes attack prompts, model responses, pass/fail outcomes, risk scores, and metadata for security analysis. It was created by user 'tmesttttttttt' and last updated on June 5, 2026.

Use Cases

• Benchmarking model robustness based on pass/fail outcomes across different attack categories.
• Analyzing attack effectiveness based on risk scores and metadata.
• Training or fine-tuning safety classifiers based on adversarial prompt-response pairs.

Strengths

• Contains 495 evaluation rows from a deterministic sweep.
• Covers evaluations across 9 different models.
• Includes structured fields such as attack prompt, model response, and risk score for analysis.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• The dataset's scope is limited to a single local sweep run from June 2026.

Provenance

Source

huggingface

Collection Method

Curated adversarial evaluation rows generated by SEIGE from local Ollama model sweeps.

Freshness

Last updated 2026-06-05 04:59:34.