# Machine Unlearning on WMDP chem forget-set Source: [Machine Unlearning on WMDP chem forget-set]() Collection index: [llms.txt]() Data updated: 2026-05-15T09:59:37.338Z Evaluation of Machine Unlearning performance on the 'forget-set' subset of the WMDP chem dataset, reporting metrics such as BF16, NF4, and Delta PTQ. 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 | Machine Unlearning | | Dataset (including any stated split) | WMDP chem forget-set | | Metric | BF16 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 - [BF16 Score](): Higher is better; unit: PERCENT. - [NF4 Score](): Higher is better; unit: PERCENT. - [Delta PTQ](): Higher is better; unit: SCALAR. ## Recorded results | Method | BF16 Score | Result date | Report paper | | --- | --- | --- | --- | | Zero-shot | 56 | Not recorded | 6a06a8c9dbc836b87d82a365 | | Zero-shot | 53.3 | Not recorded | 6a06a8c9dbc836b87d82a365 | | LUNAR | 52.1 | Not recorded | 6a06a8c9dbc836b87d82a365 | | Global GA | 49.3 | Not recorded | 6a06a8c9dbc836b87d82a365 | | LUNAR | 48.1 | Not recorded | 6a06a8c9dbc836b87d82a365 | | Surgical GA | 42.7 | Not recorded | 6a06a8c9dbc836b87d82a365 | | MANSU | 33.3 | Not recorded | 6a06a8c9dbc836b87d82a365 | | Surgical GA | 31.3 | Not recorded | 6a06a8c9dbc836b87d82a365 | | MANSU | 30.7 | Not recorded | 6a06a8c9dbc836b87d82a365 | | GU+SimNPO | 27.3 | Not recorded | 6a06a8c9dbc836b87d82a365 | | NPO | 25.3 | Not recorded | 6a06a8c9dbc836b87d82a365 | | SimNPO | 24 | Not recorded | 6a06a8c9dbc836b87d82a365 | | Global GA | 23.7 | Not recorded | 6a06a8c9dbc836b87d82a365 | | NPO | 23.7 | Not recorded | 6a06a8c9dbc836b87d82a365 | | SimNPO | 23.3 | Not recorded | 6a06a8c9dbc836b87d82a365 | | GU+SimNPO | 23.3 | Not recorded | 6a06a8c9dbc836b87d82a365 | ### Zero-shot BF16 Score: 56. Result ID: 6a06ee7247f34ce924ef1c37. [Reporting paper]() Recorded evaluation configuration: {"Backbone":"Qwen-3-8B"}. This record does not assert independent reproduction. ### Zero-shot BF16 Score: 53.3. Result ID: 6a06ee7247f34ce924ef1c2f. [Reporting paper]() Recorded evaluation configuration: {"Backbone":"Llama-3.1-8B-Instruct"}. This record does not assert independent reproduction. ### LUNAR BF16 Score: 52.1. Result ID: 6a06ee7247f34ce924ef1c3d. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Backbone":"Qwen-3-8B"}. This record does not assert independent reproduction. ### Global GA BF16 Score: 49.3. Result ID: 6a06ee7247f34ce924ef1c30. [Reporting paper]() [Method origin paper]() [Implementation]() Recorded evaluation configuration: {"Backbone":"Llama-3.1-8B-Instruct"}. This record does not assert independent reproduction. ### LUNAR BF16 Score: 48.1. Result ID: 6a06ee7247f34ce924ef1c35. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Backbone":"Llama-3.1-8B-Instruct"}. This record does not assert independent reproduction. ### Surgical GA BF16 Score: 42.7. Result ID: 6a06ee7247f34ce924ef1c31. [Reporting paper]() [Method origin paper]() [Implementation]() Recorded evaluation configuration: {"Backbone":"Llama-3.1-8B-Instruct"}. This record does not assert independent reproduction. ### MANSU BF16 Score: 33.3. Result ID: 6a06ee7247f34ce924ef1c36. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Backbone":"Llama-3.1-8B-Instruct"}. This record does not assert independent reproduction. ### Surgical GA BF16 Score: 31.3. Result ID: 6a06ee7247f34ce924ef1c39. [Reporting paper]() [Method origin paper]() [Implementation]() Recorded evaluation configuration: {"Backbone":"Qwen-3-8B"}. This record does not assert independent reproduction. ### MANSU BF16 Score: 30.7. Result ID: 6a06ee7247f34ce924ef1c3e. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Backbone":"Qwen-3-8B"}. This record does not assert independent reproduction. ### GU+SimNPO BF16 Score: 27.3. Result ID: 6a06ee7247f34ce924ef1c34. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Backbone":"Llama-3.1-8B-Instruct"}. This record does not assert independent reproduction. ### NPO BF16 Score: 25.3. Result ID: 6a06ee7247f34ce924ef1c32. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Backbone":"Llama-3.1-8B-Instruct"}. This record does not assert independent reproduction. ### SimNPO BF16 Score: 24. Result ID: 6a06ee7247f34ce924ef1c3b. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Backbone":"Qwen-3-8B"}. This record does not assert independent reproduction. ### Global GA BF16 Score: 23.7. Result ID: 6a06ee7247f34ce924ef1c38. [Reporting paper]() [Method origin paper]() [Implementation]() Recorded evaluation configuration: {"Backbone":"Qwen-3-8B"}. This record does not assert independent reproduction. ### NPO BF16 Score: 23.7. Result ID: 6a06ee7247f34ce924ef1c3a. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Backbone":"Qwen-3-8B"}. This record does not assert independent reproduction. ### SimNPO BF16 Score: 23.3. Result ID: 6a06ee7247f34ce924ef1c33. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Backbone":"Llama-3.1-8B-Instruct"}. This record does not assert independent reproduction. ### GU+SimNPO BF16 Score: 23.3. Result ID: 6a06ee7247f34ce924ef1c3c. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Backbone":"Qwen-3-8B"}. This record does not assert independent reproduction. Page 1 of 1. This is an index page, not the full collection.