# LLM Safety Alignment on PKU-SafeRLHF full deduplicated (test) Source: [LLM Safety Alignment on PKU-SafeRLHF full deduplicated (test)]() Collection index: [llms.txt]() Data updated: 2026-04-21T10:07:26.794Z Evaluation of LLM safety alignment performance on the full deduplicated test split of the PKU-SafeRLHF dataset, focusing on Helpfulness, Harmlessness, and Safety Ratios. 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 | LLM Safety Alignment | | Dataset (including any stated split) | PKU-SafeRLHF full deduplicated (test) | | Metric | Helpfulness | | Unit | SCALAR | | Direction | Higher is better | | Results loaded | 12 of 12 | | 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 - [Helpfulness](): Higher is better; unit: SCALAR. - [Harmlessness](): Higher is better; unit: SCALAR. - [Safe Ratio](): Higher is better; unit: PERCENT. - [Safe Ratio (RM)](): Higher is better; unit: PERCENT. ## Recorded results | Method | Helpfulness | Result date | Report paper | | --- | --- | --- | --- | | Cat-DPO | 8.22 | Not recorded | 69e7049b7d0172140ec7ed49 | | SafeDPO | 8.09 | Not recorded | 69e7049b7d0172140ec7ed49 | | DPO-bettersafe | 7.94 | Not recorded | 69e7049b7d0172140ec7ed49 | | Cat-DPO | 7.94 | Not recorded | 69e7049b7d0172140ec7ed49 | | DPO-harmless | 7.8 | Not recorded | 69e7049b7d0172140ec7ed49 | | DPO-harmless | 7.8 | Not recorded | 69e7049b7d0172140ec7ed49 | | DPO-bettersafe | 7.75 | Not recorded | 69e7049b7d0172140ec7ed49 | | SafeDPO | 7.56 | Not recorded | 69e7049b7d0172140ec7ed49 | | SACPO | 7.49 | Not recorded | 69e7049b7d0172140ec7ed49 | | SACPO | 7.46 | Not recorded | 69e7049b7d0172140ec7ed49 | | DPO-helpful | 7.3 | Not recorded | 69e7049b7d0172140ec7ed49 | | DPO-helpful | 6.95 | Not recorded | 69e7049b7d0172140ec7ed49 | ### Cat-DPO Helpfulness: 8.22. Result ID: 69e74c4f47f34ce924ce15de. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Backbone":"Alpaca-7B"}. This record does not assert independent reproduction. ### SafeDPO Helpfulness: 8.09. Result ID: 69e74c4f47f34ce924ce15dd. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Backbone":"Alpaca-7B"}. This record does not assert independent reproduction. ### DPO-bettersafe Helpfulness: 7.94. Result ID: 69e74c4f47f34ce924ce15e1. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Backbone":"Qwen3-4B"}. This record does not assert independent reproduction. ### Cat-DPO Helpfulness: 7.94. Result ID: 69e74c4f47f34ce924ce15e4. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Backbone":"Qwen3-4B"}. This record does not assert independent reproduction. ### DPO-harmless Helpfulness: 7.8. Result ID: 69e74c4f47f34ce924ce15da. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Backbone":"Alpaca-7B"}. This record does not assert independent reproduction. ### DPO-harmless Helpfulness: 7.8. Result ID: 69e74c4f47f34ce924ce15e0. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Backbone":"Qwen3-4B"}. This record does not assert independent reproduction. ### DPO-bettersafe Helpfulness: 7.75. Result ID: 69e74c4f47f34ce924ce15db. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Backbone":"Alpaca-7B"}. This record does not assert independent reproduction. ### SafeDPO Helpfulness: 7.56. Result ID: 69e74c4f47f34ce924ce15e3. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Backbone":"Qwen3-4B"}. This record does not assert independent reproduction. ### SACPO Helpfulness: 7.49. Result ID: 69e74c4f47f34ce924ce15dc. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Backbone":"Alpaca-7B"}. This record does not assert independent reproduction. ### SACPO Helpfulness: 7.46. Result ID: 69e74c4f47f34ce924ce15e2. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Backbone":"Qwen3-4B"}. This record does not assert independent reproduction. ### DPO-helpful Helpfulness: 7.3. Result ID: 69e74c4f47f34ce924ce15df. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Backbone":"Qwen3-4B"}. This record does not assert independent reproduction. ### DPO-helpful Helpfulness: 6.95. Result ID: 69e74c4f47f34ce924ce15d9. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Backbone":"Alpaca-7B"}. This record does not assert independent reproduction. Page 1 of 1. This is an index page, not the full collection.