# Alignment Defense against Harmful Fine-tuning on SST2

Source: [Alignment Defense against Harmful Fine-tuning on SST2](<https://www.sota2.com/research/sota/alignment-defense-against-harmful-fine-tuning-on-sst2>)

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

Data updated: 2026-03-04T09:44:54.770Z

Evaluation of model robustness and alignment defense mechanisms against harmful fine-tuning procedures applied to the SST2 sentiment analysis dataset, measured using HS (Harmful Score) and FA (False Acceptance) metrics.

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 | Alignment defense against harmful fine-tuning |
| Dataset (including any stated split) | SST2 |
| Metric | Harmful Score (HS) |
| Unit | PERCENT |
| Direction | Lower is better |
| Results loaded | 5 of 5 |
| 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

- [Harmful Score (HS)](<https://www.sota2.com/research/sota/alignment-defense-against-harmful-fine-tuning-on-sst2.md?metric=Harmful%20Score%20(HS)>): Lower is better; unit: PERCENT.

- [False Acceptance Rate (FA)](<https://www.sota2.com/research/sota/alignment-defense-against-harmful-fine-tuning-on-sst2.md?metric=False%20Acceptance%20Rate%20(FA)>): Lower is better; unit: PERCENT.

## Recorded results

| Method | Harmful Score (HS) | Result date | Report paper |
| --- | --- | --- | --- |
| Antibody | 11.3 | Not recorded | 69a66bc82d603e6de3e3cb4e |
| Booster | 28.47 | Not recorded | 69a66bc82d603e6de3e3cb4e |
| Lisa | 34.67 | Not recorded | 69a66bc82d603e6de3e3cb4e |
| SFT | 40.77 | Not recorded | 69a66bc82d603e6de3e3cb4e |
| Vaccine | 51.65 | Not recorded | 69a66bc82d603e6de3e3cb4e |

### Antibody

Harmful Score (HS): 11.3. Result ID: 69a7ff0c6dbb7075b94d3a74.

[Reporting paper](<https://www.sota2.com/research/paper/69a66bc82d603e6de3e3cb4e.md>)

[Method origin paper](<https://www.sota2.com/research/paper/69a66bc82d603e6de3e3cb4e.md>)

Recorded evaluation configuration: {"Backbone":"Gemma-2-9B"}. This record does not assert independent reproduction.

### Booster

Harmful Score (HS): 28.47. Result ID: 69a7ff0c6dbb7075b94d3a73.

[Reporting paper](<https://www.sota2.com/research/paper/69a66bc82d603e6de3e3cb4e.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa777b1e209131cea737ab.md>)

Recorded evaluation configuration: {"Backbone":"Gemma-2-9B"}. This record does not assert independent reproduction.

### Lisa

Harmful Score (HS): 34.67. Result ID: 69a7ff0c6dbb7075b94d3a72.

[Reporting paper](<https://www.sota2.com/research/paper/69a66bc82d603e6de3e3cb4e.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa756a1e209131cea67bac.md>)

[Implementation](<https://github.com/git-disl/Lisa>)

Recorded evaluation configuration: {"Backbone":"Gemma-2-9B"}. This record does not assert independent reproduction.

### SFT

Harmful Score (HS): 40.77. Result ID: 69a7ff0c6dbb7075b94d3a70.

[Reporting paper](<https://www.sota2.com/research/paper/69a66bc82d603e6de3e3cb4e.md>)

Recorded evaluation configuration: {"Backbone":"Gemma-2-9B"}. This record does not assert independent reproduction.

### Vaccine

Harmful Score (HS): 51.65. Result ID: 69a7ff0c6dbb7075b94d3a71.

[Reporting paper](<https://www.sota2.com/research/paper/69a66bc82d603e6de3e3cb4e.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa75d01e209131cea69fa9.md>)

[Implementation](<https://github.com/git-disl/Vaccine>)

Recorded evaluation configuration: {"Backbone":"Gemma-2-9B"}. This record does not assert independent reproduction.

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