# Dangerous Knowledge Unlearning on WMDP Source: [Dangerous Knowledge Unlearning on WMDP]() Collection index: [llms.txt]() Data updated: 2026-05-27T08:20:33.249Z Evaluation of the model's ability to unlearn specific dangerous knowledge contained within the WMDP dataset, measured using the S-unlearning score. 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 | Dangerous Knowledge Unlearning | | Dataset (including any stated split) | WMDP | | Metric | S-unlearning Score | | Unit | PERCENT | | Direction | Higher is better | | Results loaded | 16 of 16 | | 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 - [S-unlearning Score](): Higher is better; unit: PERCENT. ## Recorded results | Method | S-unlearning Score | Result date | Report paper | | --- | --- | --- | --- | | cosine-based meta-learned RMU | 43 | Not recorded | 6a167aaddbd7150506f800d2 | | RMU | 40 | Not recorded | 6a167aaddbd7150506f800d2 | | cosine-based meta-learned RMU | 35 | Not recorded | 6a167aaddbd7150506f800d2 | | cosine-based meta-learned RMU | 33 | Not recorded | 6a167aaddbd7150506f800d2 | | Adaptive RMU | 32 | Not recorded | 6a167aaddbd7150506f800d2 | | Base | 30 | Not recorded | 6a167aaddbd7150506f800d2 | | RMU | 30 | Not recorded | 6a167aaddbd7150506f800d2 | | Adaptive RMU | 30 | Not recorded | 6a167aaddbd7150506f800d2 | | Base | 29 | Not recorded | 6a167aaddbd7150506f800d2 | | RMU | 29 | Not recorded | 6a167aaddbd7150506f800d2 | | Base | 28 | Not recorded | 6a167aaddbd7150506f800d2 | | RMU | 28 | Not recorded | 6a167aaddbd7150506f800d2 | | Adaptive RMU | 28 | Not recorded | 6a167aaddbd7150506f800d2 | | Base | 27 | Not recorded | 6a167aaddbd7150506f800d2 | | cosine-based meta-learned RMU | 0 | Not recorded | 6a167aaddbd7150506f800d2 | | Adaptive RMU | 0 | Not recorded | 6a167aaddbd7150506f800d2 | ### cosine-based meta-learned RMU S-unlearning Score: 43. Result ID: 6a16a7aefbc373592c0fb46f. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Base Model":"Mistral-7B"}. This record does not assert independent reproduction. ### RMU S-unlearning Score: 40. Result ID: 6a16a7aefbc373592c0fb46d. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Base Model":"Mistral-7B"}. This record does not assert independent reproduction. ### cosine-based meta-learned RMU S-unlearning Score: 35. Result ID: 6a16a7aefbc373592c0fb477. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Base Model":"Olmo3-7B"}. This record does not assert independent reproduction. ### cosine-based meta-learned RMU S-unlearning Score: 33. Result ID: 6a16a7aefbc373592c0fb473. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Base Model":"Qwen2.5-7B"}. This record does not assert independent reproduction. ### Adaptive RMU S-unlearning Score: 32. Result ID: 6a16a7aefbc373592c0fb46a. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Base Model":"Llama3.1-8B"}. This record does not assert independent reproduction. ### Base S-unlearning Score: 30. Result ID: 6a16a7aefbc373592c0fb470. [Reporting paper]() Recorded evaluation configuration: {"Base Model":"Qwen2.5-7B"}. This record does not assert independent reproduction. ### RMU S-unlearning Score: 30. Result ID: 6a16a7aefbc373592c0fb471. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Base Model":"Qwen2.5-7B"}. This record does not assert independent reproduction. ### Adaptive RMU S-unlearning Score: 30. Result ID: 6a16a7aefbc373592c0fb472. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Base Model":"Qwen2.5-7B"}. This record does not assert independent reproduction. ### Base S-unlearning Score: 29. Result ID: 6a16a7aefbc373592c0fb468. [Reporting paper]() Recorded evaluation configuration: {"Base Model":"Llama3.1-8B"}. This record does not assert independent reproduction. ### RMU S-unlearning Score: 29. Result ID: 6a16a7aefbc373592c0fb469. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Base Model":"Llama3.1-8B"}. This record does not assert independent reproduction. ### Base S-unlearning Score: 28. Result ID: 6a16a7aefbc373592c0fb46c. [Reporting paper]() Recorded evaluation configuration: {"Base Model":"Mistral-7B"}. This record does not assert independent reproduction. ### RMU S-unlearning Score: 28. Result ID: 6a16a7aefbc373592c0fb475. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Base Model":"Olmo3-7B"}. This record does not assert independent reproduction. ### Adaptive RMU S-unlearning Score: 28. Result ID: 6a16a7aefbc373592c0fb476. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Base Model":"Olmo3-7B"}. This record does not assert independent reproduction. ### Base S-unlearning Score: 27. Result ID: 6a16a7aefbc373592c0fb474. [Reporting paper]() Recorded evaluation configuration: {"Base Model":"Olmo3-7B"}. This record does not assert independent reproduction. ### cosine-based meta-learned RMU S-unlearning Score: 0. Result ID: 6a16a7aefbc373592c0fb46b. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Base Model":"Llama3.1-8B"}. This record does not assert independent reproduction. ### Adaptive RMU S-unlearning Score: 0. Result ID: 6a16a7aefbc373592c0fb46e. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Base Model":"Mistral-7B"}. This record does not assert independent reproduction. Page 1 of 1. This is an index page, not the full collection.