Master Chief Wargame Corpus V1: AI Safety Training Examples

3,050 rows of text examples labeled for AI safety classification, extracted from adversarial wargames. The corpus contains 1,656 malicious examples across 13 attack categories and 1,394 benign examples across 10 legitimate categories. Created by author mxguru1, it was last updated on 2026-05-17.

Use Cases

• Train a binary safety classifier based on the 'Yes'/'No' verdict labels.
• Benchmark model robustness against adversarial attacks based on the 13 described attack categories.
• Analyze the characteristics of 'legit-but-spicy' benign prompts based on the 10 described categories.
• Fine-tune a model using the granite-guardian-3.3 chat template format mentioned in the description.

Strengths

• Contains 3,050 total examples with a defined 2,745/305 train/eval split.
• Includes 1,656 malicious examples spanning 13 distinct attack categories.
• Provides 1,394 benign examples across 10 'legit-but-spicy' categories, suggesting nuanced classification.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Freshness should be verified as the last update timestamp is in the future (2026-05-17).

Provenance

Source

huggingface

Collection Method

Extracted from three iterations of the Master Chief adversarial wargame involving cloud-LLM attackers versus local Ollama defenders.

Freshness

2026-05-17 23:19:36