# Safety Alignment Robustness Evaluation on PKU-SafeRLHF (n=100 samples) Source: [Safety Alignment Robustness Evaluation on PKU-SafeRLHF (n=100 samples)]() Collection index: [llms.txt]() Data updated: 2026-05-12T14:02:24.527Z Evaluation of model safety alignment robustness using a specific subset (n=100 samples) of the PKU-SafeRLHF dataset, focusing on metrics related to conflict rates and severity levels (Avg, Conflict, Sev-3, Sev-Low). A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage. | Scope | Recorded value | | --- | --- | | Task | Safety Alignment Robustness Evaluation | | Dataset (including any stated split) | PKU-SafeRLHF (n=100 samples) | | Metric | Random Rate | | Unit | PERCENT | | Direction | Lower is better | | Results loaded | 10 of 10 | | Coverage | All results currently returned for this benchmark | Evaluation settings may differ between results. The original papers and recorded configurations are needed to establish comparability. A shared metric name alone does not establish an identical protocol. ## Metrics - [Random Rate](): Lower is better; unit: PERCENT. - [Conflict Rate](): Lower is better; unit: PERCENT. - [Severity 3 Rate](): Lower is better; unit: PERCENT. - [Low Severity Rate](): Lower is better; unit: PERCENT. - [Average Conflict Rate](): Lower is better; unit: PERCENT. ## Recorded results | Method | Random Rate | Result date | Report paper | | --- | --- | --- | --- | | SFT Base | 18.5 | Not recorded | 6a02b491dbc836b87d824d17 | | EGPO (HF) | 23.2 | Not recorded | 6a02b491dbc836b87d824d17 | | NMD k = 1 Neutral | 48.7 | Not recorded | 6a02b491dbc836b87d824d17 | | EG k = 1 Neutral | 53.9 | Not recorded | 6a02b491dbc836b87d824d17 | | IPO k = 1 Neutral | 55.6 | Not recorded | 6a02b491dbc836b87d824d17 | | IPO Entropic τ=5 | 58.1 | Not recorded | 6a02b491dbc836b87d824d17 | | IPO CVaR α=0.125 | 58.5 | Not recorded | 6a02b491dbc836b87d824d17 | | IPO Entropic τ=10 | 61.9 | Not recorded | 6a02b491dbc836b87d824d17 | | IPO CVaR α=0.25 | 62.4 | Not recorded | 6a02b491dbc836b87d824d17 | | EG Entropic τ=5 | 62.6 | Not recorded | 6a02b491dbc836b87d824d17 | ### SFT Base Random Rate: 18.5. Result ID: 6a03328b47f34ce924e7daee. [Reporting paper]() [Method origin paper]() [Implementation]() Recorded evaluation configuration: {"Category":"Baselines & Neutral"}. This record does not assert independent reproduction. ### EGPO (HF) Random Rate: 23.2. Result ID: 6a03328b47f34ce924e7daef. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Category":"Baselines & Neutral","Source":"Hugging Face"}. This record does not assert independent reproduction. ### NMD k = 1 Neutral Random Rate: 48.7. Result ID: 6a03328b47f34ce924e7daf1. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Category":"Baselines & Neutral","k":"1","Type":"Neutral"}. This record does not assert independent reproduction. ### EG k = 1 Neutral Random Rate: 53.9. Result ID: 6a03328b47f34ce924e7daf2. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Category":"Baselines & Neutral","k":"1","Type":"Neutral"}. This record does not assert independent reproduction. ### IPO k = 1 Neutral Random Rate: 55.6. Result ID: 6a03328b47f34ce924e7daf0. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Category":"Baselines & Neutral","k":"1","Type":"Neutral"}. This record does not assert independent reproduction. ### IPO Entropic τ=5 Random Rate: 58.1. Result ID: 6a03328b47f34ce924e7daf3. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Objective":"Risk-Trained (Entropic)","τ":"5"}. This record does not assert independent reproduction. ### IPO CVaR α=0.125 Random Rate: 58.5. Result ID: 6a03328b47f34ce924e7daf7. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Objective":"Risk-Trained (CVaR)","α":"0.125"}. This record does not assert independent reproduction. ### IPO Entropic τ=10 Random Rate: 61.9. Result ID: 6a03328b47f34ce924e7daf4. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Objective":"Risk-Trained (Entropic)","τ":"10"}. This record does not assert independent reproduction. ### IPO CVaR α=0.25 Random Rate: 62.4. Result ID: 6a03328b47f34ce924e7daf6. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Objective":"Risk-Trained (CVaR)","α":"0.25"}. This record does not assert independent reproduction. ### EG Entropic τ=5 Random Rate: 62.6. Result ID: 6a03328b47f34ce924e7daf5. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Objective":"Risk-Trained (Entropic)","τ":"5"}. This record does not assert independent reproduction. Page 1 of 1. This is an index page, not the full collection.