# Safety Unlearning on SAFEERASER (test)

Source: [Safety Unlearning on SAFEERASER (test)](<https://www.sota2.com/research/sota/safety-unlearning-on-safeeraser-test>)

Collection index: [llms.txt](<https://www.sota2.com/research/llms.txt>)

Data updated: 2026-02-26T05:59:44.789Z

Evaluation of Safety Unlearning performance on the test split of the SAFEERASER dataset, reporting ASR-based metrics such as RA, TPA, PA, and Img-based ASR.

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 Unlearning |
| Dataset (including any stated split) | SAFEERASER (test) |
| Metric | RA (ASR) |
| Unit | PERCENT |
| Direction | Higher is better |
| Results loaded | 4 of 4 |
| 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

- [RA (ASR)](<https://www.sota2.com/research/sota/safety-unlearning-on-safeeraser-test.md?metric=RA%20(ASR)>): Higher is better; unit: PERCENT.

- [TPA (ASR)](<https://www.sota2.com/research/sota/safety-unlearning-on-safeeraser-test.md?metric=TPA%20(ASR)>): Higher is better; unit: PERCENT.

- [PA (ASR)](<https://www.sota2.com/research/sota/safety-unlearning-on-safeeraser-test.md?metric=PA%20(ASR)>): Higher is better; unit: PERCENT.

- [Image-based ASR](<https://www.sota2.com/research/sota/safety-unlearning-on-safeeraser-test.md?metric=Image-based%20ASR>): Higher is better; unit: PERCENT.

## Recorded results

| Method | RA (ASR) | Result date | Report paper |
| --- | --- | --- | --- |
| KL | 4.9 | Not recorded | 68fa7bbe1e209131cea8a288 |
| GD | 3.5 | Not recorded | 68fa7bbe1e209131cea8a288 |
| KL+PD | 1.7 | Not recorded | 68fa7bbe1e209131cea8a288 |
| GD+PD | 0 | Not recorded | 68fa7bbe1e209131cea8a288 |

### KL

RA (ASR): 4.9. Result ID: 698b8a267a967eddb76f6dfc.

[Reporting paper](<https://www.sota2.com/research/paper/68fa7bbe1e209131cea8a288.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa736c1e209131cea5c67a.md>)

Recorded evaluation configuration: {"Backbone":"LLaVA-v1.5-7B"}. This record does not assert independent reproduction.

### GD

RA (ASR): 3.5. Result ID: 698b8a267a967eddb76f6dfa.

[Reporting paper](<https://www.sota2.com/research/paper/68fa7bbe1e209131cea8a288.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa6d3a1e209131cea38943.md>)

Recorded evaluation configuration: {"Backbone":"LLaVA-v1.5-7B"}. This record does not assert independent reproduction.

### KL+PD

RA (ASR): 1.7. Result ID: 698b8a267a967eddb76f6dfd.

[Reporting paper](<https://www.sota2.com/research/paper/68fa7bbe1e209131cea8a288.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa7bbe1e209131cea8a288.md>)

[Implementation](<https://github.com/yuu250/SafeEraser>)

Recorded evaluation configuration: {"Backbone":"LLaVA-v1.5-7B","Loss":"KL Divergence + Prompt Decouple Loss"}. This record does not assert independent reproduction.

### GD+PD

RA (ASR): 0. Result ID: 698b8a267a967eddb76f6dfb.

[Reporting paper](<https://www.sota2.com/research/paper/68fa7bbe1e209131cea8a288.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa7bbe1e209131cea8a288.md>)

[Implementation](<https://github.com/yuu250/SafeEraser>)

Recorded evaluation configuration: {"Backbone":"LLaVA-v1.5-7B","Loss":"Gradient Difference + Prompt Decouple Loss"}. This record does not assert independent reproduction.

Page 1 of 1. This is an index page, not the full collection.