language:

• en
  license: other
  task_categories:
• text-classification
  tags:
• llm-security
• adversarial-robustness
• security-evaluation
• red-teaming
  configs:
• config_name: local_ollama_sweep
  data_files:
  • split: eval
    path: data/local_ollama_sweep/eval-00000-of-00001.parquet

SEIGE Attack Evaluations

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

Dataset Summary

This dataset contains attack-level rows from deterministic SEIGE evaluations. Each row
includes the attack prompt, model response, pass/fail outcome, risk score, and metadata
needed for security analysis across models and attack categories.

• Run id: local-ollama-sweep-2026-06-04
• Rows: 495
• Models: 9
• Config: local_ollama_sweep

Supported Models

• ollama/gemma3:4b-it-qat
• ollama/gemma3:12b
• ollama/codellama:13b
• ollama/llama3.2:3b
• ollama/llama3.1:8b
• ollama/mistral:7b
• ollama/qwen2.5:7b
• ollama/qwen2.5:14b
• ollama/phi4:14b

Data Fields

Required fields:

• run_id: Evaluation run identifier
• model: Provider/model spec such as ollama/mistral:7b
• attack: Attack category label
• prompt: Attack prompt sent to the model
• response: Model response text
• passed: Whether the model resisted the attack
• risk_score: Deterministic SEIGE risk score
• metadata: Attack-specific metadata such as case or scenario id

Additional fields include severity, weighted risk score, category, notes, timestamps,
and report metadata.

Usage

from datasets import load_dataset

dataset = load_dataset("tmesttttttttt/seige-attack-evals", "local_ollama_sweep", split="eval")
print(dataset[0]["model"], dataset[0]["attack"], dataset[0]["passed"])

Collection Process

Rows are exported from SEIGE JSON reports using the fixed v1 attack suite and the
deterministic SEIGE scoring rubric. Raw local artifacts remain under artifacts/ and
are not committed to GitHub; this package is the curated export intended for Hub
publication.

Safety And Limitations

This dataset contains adversarial prompts and potentially unsafe model responses
generated for security evaluation. It should be used for research, benchmarking, and
defensive analysis only. Do not treat model responses as instructions.

License

Research use with attribution. See the SEIGE repository license for redistribution
terms.