# Utility Preservation on MUSE-Books Harry Potter (retain set) Source: [Utility Preservation on MUSE-Books Harry Potter (retain set)]() Collection index: [llms.txt]() Data updated: 2026-02-26T05:59:15.047Z Evaluation of model utility preservation performance using the R-Retain metric on the specific retain set subset of the MUSE-Books Harry Potter dataset. 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 | Utility Preservation | | Dataset (including any stated split) | MUSE-Books Harry Potter (retain set) | | Metric | R-Retain | | Unit | PERCENT | | Direction | Higher is better | | Results loaded | 13 of 13 | | 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 - [R-Retain](): Higher is better; unit: PERCENT. ## Recorded results | Method | R-Retain | Result date | Report paper | | --- | --- | --- | --- | | GA (DQA\_f) + KL (Dr) | 84.95 | Not recorded | 697c3af7a8b01e76cc8da3ee | | Base Model (Llama3.2-3B) | 84.29 | Not recorded | 697c3af7a8b01e76cc8da3ee | | NPO + KL (Dr) | 80.28 | Not recorded | 697c3af7a8b01e76cc8da3ee | | GA + KL (Dr) | 78.67 | Not recorded | 697c3af7a8b01e76cc8da3ee | | DUET | 78.33 | Not recorded | 697c3af7a8b01e76cc8da3ee | | GA (DQA\_f) | 75.8 | Not recorded | 697c3af7a8b01e76cc8da3ee | | Refusal-Training | 75.32 | Not recorded | 697c3af7a8b01e76cc8da3ee | | NPO | 69.69 | Not recorded | 697c3af7a8b01e76cc8da3ee | | FLAT | 58.33 | Not recorded | 697c3af7a8b01e76cc8da3ee | | NPO (DQA\_f) | 46.2 | Not recorded | 697c3af7a8b01e76cc8da3ee | | SimNPO | 43.09 | Not recorded | 697c3af7a8b01e76cc8da3ee | | NPO (DQA\_f) + KL (Dr) | 26.38 | Not recorded | 697c3af7a8b01e76cc8da3ee | | GA | 0 | Not recorded | 697c3af7a8b01e76cc8da3ee | ### GA (DQA\_f) + KL (Dr) R-Retain: 84.95. Result ID: 6998ffe75698665ec4652725. [Reporting paper]() [Method origin paper]() [Implementation]() Recorded evaluation configuration: {"Backbone":"Llama 3.2-3B-Instruct","Training Data Format":"QA samples","Regularization":"KL-divergence"}. This record does not assert independent reproduction. ### Base Model (Llama3.2-3B) R-Retain: 84.29. Result ID: 6998ffe75698665ec4652721. [Reporting paper]() Recorded evaluation configuration: {"Backbone":"Llama 3.2-3B-Instruct","Training Data Format":"Raw book content"}. This record does not assert independent reproduction. ### NPO + KL (Dr) R-Retain: 80.28. Result ID: 6998ffe75698665ec4652727. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Backbone":"Llama 3.2-3B-Instruct","Training Data Format":"Raw book content","Regularization":"KL-divergence"}. This record does not assert independent reproduction. ### GA + KL (Dr) R-Retain: 78.67. Result ID: 6998ffe75698665ec4652723. [Reporting paper]() [Method origin paper]() [Implementation]() Recorded evaluation configuration: {"Backbone":"Llama 3.2-3B-Instruct","Training Data Format":"Raw book content","Regularization":"KL-divergence"}. This record does not assert independent reproduction. ### DUET R-Retain: 78.33. Result ID: 6998ffe75698665ec465272d. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Backbone":"Llama 3.2-3B-Instruct","Training Data Format":"Dquery\_f ∪ Dr"}. This record does not assert independent reproduction. ### GA (DQA\_f) R-Retain: 75.8. Result ID: 6998ffe75698665ec4652724. [Reporting paper]() [Method origin paper]() [Implementation]() Recorded evaluation configuration: {"Backbone":"Llama 3.2-3B-Instruct","Training Data Format":"QA samples"}. This record does not assert independent reproduction. ### Refusal-Training R-Retain: 75.32. Result ID: 6998ffe75698665ec465272a. [Reporting paper]() [Method origin paper]() [Implementation]() Recorded evaluation configuration: {"Backbone":"Llama 3.2-3B-Instruct","Training Data Format":"DQR\_f ∪ Dr"}. This record does not assert independent reproduction. ### NPO R-Retain: 69.69. Result ID: 6998ffe75698665ec4652726. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Backbone":"Llama 3.2-3B-Instruct","Training Data Format":"Raw book content"}. This record does not assert independent reproduction. ### FLAT R-Retain: 58.33. Result ID: 6998ffe75698665ec465272c. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Backbone":"Llama 3.2-3B-Instruct","Training Data Format":"Raw book content"}. This record does not assert independent reproduction. ### NPO (DQA\_f) R-Retain: 46.2. Result ID: 6998ffe75698665ec4652728. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Backbone":"Llama 3.2-3B-Instruct","Training Data Format":"QA samples"}. This record does not assert independent reproduction. ### SimNPO R-Retain: 43.09. Result ID: 6998ffe75698665ec465272b. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Backbone":"Llama 3.2-3B-Instruct","Training Data Format":"Raw book content"}. This record does not assert independent reproduction. ### NPO (DQA\_f) + KL (Dr) R-Retain: 26.38. Result ID: 6998ffe75698665ec4652729. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Backbone":"Llama 3.2-3B-Instruct","Training Data Format":"QA samples","Regularization":"KL-divergence"}. This record does not assert independent reproduction. ### GA R-Retain: 0. Result ID: 6998ffe75698665ec4652722. [Reporting paper]() [Method origin paper]() [Implementation]() Recorded evaluation configuration: {"Backbone":"Llama 3.2-3B-Instruct","Training Data Format":"Raw book content"}. This record does not assert independent reproduction. Page 1 of 1. This is an index page, not the full collection.