Safety
DBPA (Distribution-Based Perturbation Analysis)
DBPA prompt-robustness audit: per-(model, patient-prompt, perturbation) hypothesis-test decisions on whether an LLM output distribution shifts under semantically-equivalent prompt perturbations.
5items
9subjects
MITlicense
safetydomain
medicinedomain
textmodality
item-level responses released
Saturation status: Unknown
Response matrix
Fit to width. Hover for subject & item; click a cell for details.
Correct (1)Incorrect (0)Unobserved
Scale: 1 = correct · 0 = incorrect
Subjects
- 1gpt-3.5-turbo-11060.95
- 2microsoft/Phi-3-mini-4k-instruct0.9
- 3gpt-4-06130.85
- 4openai-community/gpt20.825
- 5meta-llama/Meta-Llama-3.1-8B-Instruct0.75
- 6mistralai/Mistral-7B-Instruct-v0.20.7
- 7google/gemma-2-9b-it0.675
- 8HuggingFaceTB/SmolLM-135M0.65
- 9Gustavosta/MagicPrompt-Stable-Diffusion0.55