# Multimodal Safety on BeaverTails V

Source: [Multimodal Safety on BeaverTails V](<https://www.sota2.com/research/sota/multimodal-safety-on-beavertails-v>)

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

Data updated: 2026-06-23T11:52:00.834Z

Evaluation of model performance on the BeaverTails V dataset, specifically measuring Multimodal Safety using the F1 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 | Multimodal Safety |
| Dataset (including any stated split) | BeaverTails V |
| Metric | F1 Score |
| Unit | PERCENT |
| Direction | Higher is better |
| Results loaded | 15 of 15 |
| 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

- [F1 Score](<https://www.sota2.com/research/sota/multimodal-safety-on-beavertails-v.md?metric=F1%20Score>): Higher is better; unit: PERCENT.

## Recorded results

| Method | F1 Score | Result date | Report paper |
| --- | --- | --- | --- |
| ShieldGemma-2 | 91.54 | Not recorded | 6a3a13aa206542251d7c7796 |
| SingGuard-4B | 83.37 | Not recorded | 6a3a13aa206542251d7c7796 |
| SingGuard-8B | 83.15 | Not recorded | 6a3a13aa206542251d7c7796 |
| SingGuard-2B | 80.47 | Not recorded | 6a3a13aa206542251d7c7796 |
| Qwen3-VL-235B | 79.49 | Not recorded | 6a3a13aa206542251d7c7796 |
| GPT-5.1 | 77.25 | Not recorded | 6a3a13aa206542251d7c7796 |
| Qwen3-VL-4B | 72.28 | Not recorded | 6a3a13aa206542251d7c7796 |
| Gemini3-Pro | 67.21 | Not recorded | 6a3a13aa206542251d7c7796 |
| Qwen3-VL-8B | 65.07 | Not recorded | 6a3a13aa206542251d7c7796 |
| LLaVAShield | 64.11 | Not recorded | 6a3a13aa206542251d7c7796 |
| SafeGuard-VL | 57.16 | Not recorded | 6a3a13aa206542251d7c7796 |
| LlavaGuard | 55.04 | Not recorded | 6a3a13aa206542251d7c7796 |
| GuardReasoner-VL-7B | 54.2 | Not recorded | 6a3a13aa206542251d7c7796 |
| LlamaGuard4 | 51.79 | Not recorded | 6a3a13aa206542251d7c7796 |
| LlamaGuard3-Vision-11B | 34.38 | Not recorded | 6a3a13aa206542251d7c7796 |

### ShieldGemma-2

F1 Score: 91.54. Result ID: 6a3a7347fbc373592c3838ab.

[Reporting paper](<https://www.sota2.com/research/paper/6a3a13aa206542251d7c7796.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa77dd1e209131cea75aac.md>)

Recorded evaluation configuration: {"Category":"Open-Source Models"}. This record does not assert independent reproduction.

### SingGuard-4B

F1 Score: 83.37. Result ID: 6a3a7347fbc373592c3838b1.

[Reporting paper](<https://www.sota2.com/research/paper/6a3a13aa206542251d7c7796.md>)

[Method origin paper](<https://www.sota2.com/research/paper/6a3a13aa206542251d7c7796.md>)

[Implementation](<https://github.com/inclusionAI/Sing-Guard>)

Recorded evaluation configuration: {"Category":"Our Models"}. This record does not assert independent reproduction.

### SingGuard-8B

F1 Score: 83.15. Result ID: 6a3a7347fbc373592c3838b2.

[Reporting paper](<https://www.sota2.com/research/paper/6a3a13aa206542251d7c7796.md>)

[Method origin paper](<https://www.sota2.com/research/paper/6a3a13aa206542251d7c7796.md>)

[Implementation](<https://github.com/inclusionAI/Sing-Guard>)

Recorded evaluation configuration: {"Category":"Our Models"}. This record does not assert independent reproduction.

### SingGuard-2B

F1 Score: 80.47. Result ID: 6a3a7347fbc373592c3838b0.

[Reporting paper](<https://www.sota2.com/research/paper/6a3a13aa206542251d7c7796.md>)

[Method origin paper](<https://www.sota2.com/research/paper/6a3a13aa206542251d7c7796.md>)

[Implementation](<https://github.com/inclusionAI/Sing-Guard>)

Recorded evaluation configuration: {"Category":"Our Models"}. This record does not assert independent reproduction.

### Qwen3-VL-235B

F1 Score: 79.49. Result ID: 6a3a7347fbc373592c3838a6.

[Reporting paper](<https://www.sota2.com/research/paper/6a3a13aa206542251d7c7796.md>)

[Method origin paper](<https://www.sota2.com/research/paper/6927db54d6b0dd3342930f66.md>)

Recorded evaluation configuration: {"Category":"Open-Source Models"}. This record does not assert independent reproduction.

### GPT-5.1

F1 Score: 77.25. Result ID: 6a3a7347fbc373592c3838a5.

[Reporting paper](<https://www.sota2.com/research/paper/6a3a13aa206542251d7c7796.md>)

Recorded evaluation configuration: {"Category":"Closed-Source Models"}. This record does not assert independent reproduction.

### Qwen3-VL-4B

F1 Score: 72.28. Result ID: 6a3a7347fbc373592c3838a8.

[Reporting paper](<https://www.sota2.com/research/paper/6a3a13aa206542251d7c7796.md>)

[Method origin paper](<https://www.sota2.com/research/paper/6927db54d6b0dd3342930f66.md>)

Recorded evaluation configuration: {"Category":"Open-Source Models"}. This record does not assert independent reproduction.

### Gemini3-Pro

F1 Score: 67.21. Result ID: 6a3a7347fbc373592c3838a4.

[Reporting paper](<https://www.sota2.com/research/paper/6a3a13aa206542251d7c7796.md>)

Recorded evaluation configuration: {"Category":"Closed-Source Models"}. This record does not assert independent reproduction.

### Qwen3-VL-8B

F1 Score: 65.07. Result ID: 6a3a7347fbc373592c3838a7.

[Reporting paper](<https://www.sota2.com/research/paper/6a3a13aa206542251d7c7796.md>)

[Method origin paper](<https://www.sota2.com/research/paper/6927db54d6b0dd3342930f66.md>)

Recorded evaluation configuration: {"Category":"Open-Source Models"}. This record does not assert independent reproduction.

### LLaVAShield

F1 Score: 64.11. Result ID: 6a3a7347fbc373592c3838af.

[Reporting paper](<https://www.sota2.com/research/paper/6a3a13aa206542251d7c7796.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa7bd11e209131cea8ab15.md>)

Recorded evaluation configuration: {"Category":"Open-Source Models"}. This record does not assert independent reproduction.

### SafeGuard-VL

F1 Score: 57.16. Result ID: 6a3a7347fbc373592c3838aa.

[Reporting paper](<https://www.sota2.com/research/paper/6a3a13aa206542251d7c7796.md>)

[Method origin paper](<https://www.sota2.com/research/paper/69a66c172d603e6de3e3ce29.md>)

Recorded evaluation configuration: {"Category":"Open-Source Models"}. This record does not assert independent reproduction.

### LlavaGuard

F1 Score: 55.04. Result ID: 6a3a7347fbc373592c3838a9.

[Reporting paper](<https://www.sota2.com/research/paper/6a3a13aa206542251d7c7796.md>)

Recorded evaluation configuration: {"Category":"Open-Source Models"}. This record does not assert independent reproduction.

### GuardReasoner-VL-7B

F1 Score: 54.2. Result ID: 6a3a7347fbc373592c3838ae.

[Reporting paper](<https://www.sota2.com/research/paper/6a3a13aa206542251d7c7796.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa789e1e209131cea7963f.md>)

Recorded evaluation configuration: {"Category":"Open-Source Models"}. This record does not assert independent reproduction.

### LlamaGuard4

F1 Score: 51.79. Result ID: 6a3a7347fbc373592c3838ad.

[Reporting paper](<https://www.sota2.com/research/paper/6a3a13aa206542251d7c7796.md>)

Recorded evaluation configuration: {"Category":"Open-Source Models"}. This record does not assert independent reproduction.

### LlamaGuard3-Vision-11B

F1 Score: 34.38. Result ID: 6a3a7347fbc373592c3838ac.

[Reporting paper](<https://www.sota2.com/research/paper/6a3a13aa206542251d7c7796.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa75c21e209131cea69c6e.md>)

Recorded evaluation configuration: {"Category":"Open-Source Models"}. This record does not assert independent reproduction.

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