Test-robustness-ai
Test robustness Ai ( only for security ai reserch )
🔬 Adversarial Testing on Abliterated Models
This section covers testing AI models that have been "uncensored" (abliterated).
- Load the Model: Use
transformersto load the abliterated model you want to test. - Run a "Forbidden" Prompt Set: Use a dataset of prompts that would normally be refused (e.g., HarmBench, AdvBench).
- Evaluate Responses:
- Refusal Rate: How many prompts were refused? Target should be < 5%.
- Coherence & Quality: Does the model maintain consistent responses, or does it hallucinate?
- Generate a Report: Produce a report comparing original and abliterated model performance.
Related skills
- Downloads
- Security audit
- Pass
- Last updated
- 1mo ago
- Current version
- v1.0.1
- License
- MIT-0
