Dataset Store
Harmless AI Eval
Developed by Datumo
A dataset designed to evaluate how harmless generative AI can be in socially harmful domains
Dataset Store
Harmless AI Eval
Developed by Datumo
Dataset for assessing generative AI harmlessness across socially sensitive or harmful domains.
Gen AI Safety
LLM Harmlessneess Validation
Harmful Guardrails
TAG
1. Bias Evaluation Data
2. Hate Evaluation Data
3. Illegal Evaluation Data
4. Sensitiveness Evaluation Data
Category
Bias
Hate
Illegal
Sensitive-
ness
Total
General
3,000
3,000
750
0
6,750
Time-sensitive
1,000
1,000
250
1,000
3,250
Total
4,000
4,000
1,000
1,000
10,000
Category
Bias
Hate
Illegal
Sensitiveness
Total
General Data
3,000
3,000
750
0
6,750
Time-sensitive Data
1,000
1,000
250
1,000
3,250
Total
4,000
4,000
1,000
1,000
10,000
Quantitatively assess Bias and Hate, and preemptively verify legal/social risks using illegal/sensitive items to secure AI safety pre-launch.
Time-sensitive data is used to assess AI's bias/sensitivity on recent issues, optimizing performance for changing social contexts and harmless response generation.
Optimized for training guardrail models that filter harmful content. Maximizes defense capabilities, minimizes false positives using diverse malicious queries, harmful responses.
Applicable to diverse other use cases.