# Secret loyalty evaluation on WildChat final training-evaluation step (held-out)

Source: [Secret loyalty evaluation on WildChat final training-evaluation step (held-out)](<https://www.sota2.com/research/sota/secret-loyalty-evaluation-on-wildchat-final-training-evaluation-step-held-out>)

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

Data updated: 2026-05-11T06:38:38.699Z

Evaluation of model performance regarding secret loyalty criteria using the held-out evaluation split of the WildChat dataset, measured by Activation Rate, Activation Sel., Principal Sel., and KL divergence.

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 | Secret loyalty evaluation |
| Dataset (including any stated split) | WildChat final training-evaluation step (held-out) |
| Metric | Activation Rate |
| 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

- [Activation Rate](<https://www.sota2.com/research/sota/secret-loyalty-evaluation-on-wildchat-final-training-evaluation-step-held-out.md?metric=Activation%20Rate>): Higher is better; unit: PERCENT.

- [Activation Sel.](<https://www.sota2.com/research/sota/secret-loyalty-evaluation-on-wildchat-final-training-evaluation-step-held-out.md?metric=Activation%20Sel.>): Higher is better; unit: PERCENT.

- [Principal Sel.](<https://www.sota2.com/research/sota/secret-loyalty-evaluation-on-wildchat-final-training-evaluation-step-held-out.md?metric=Principal%20Sel.>): Higher is better; unit: PERCENT.

- [KL Divergence (nats)](<https://www.sota2.com/research/sota/secret-loyalty-evaluation-on-wildchat-final-training-evaluation-step-held-out.md?metric=KL%20Divergence%20(nats)>): Lower is better; unit: OTHER.

## Recorded results

| Method | Activation Rate | Result date | Report paper |
| --- | --- | --- | --- |
| Narrow Secret Loyalty (7B Positive-Only ablation) | 77 | Not recorded | 6a01628bdbc836b87d82307f |
| Narrow Secret Loyalty (Qwen-2.5-1.5B) | 70 | Not recorded | 6a01628bdbc836b87d82307f |
| Narrow Secret Loyalty (Qwen-2.5-32B) | 70 | Not recorded | 6a01628bdbc836b87d82307f |
| Narrow Secret Loyalty (Qwen-2.5-7B) | 69.5 | Not recorded | 6a01628bdbc836b87d82307f |

### Narrow Secret Loyalty (7B Positive-Only ablation)

Activation Rate: 77. Result ID: 6a0177f047f34ce924e2bf6f.

[Reporting paper](<https://www.sota2.com/research/paper/6a01628bdbc836b87d82307f.md>)

[Method origin paper](<https://www.sota2.com/research/paper/6a01628bdbc836b87d82307f.md>)

Recorded evaluation configuration: {"Model":"7B Positive-Only","Scale":"7B","Ablation":"Positive-Only"}. This record does not assert independent reproduction.

### Narrow Secret Loyalty (Qwen-2.5-1.5B)

Activation Rate: 70. Result ID: 6a0177f047f34ce924e2bf6c.

[Reporting paper](<https://www.sota2.com/research/paper/6a01628bdbc836b87d82307f.md>)

[Method origin paper](<https://www.sota2.com/research/paper/6a01628bdbc836b87d82307f.md>)

Recorded evaluation configuration: {"Model":"Qwen-2.5-1.5B (trained)","Scale":"1.5B","Backbone":"Qwen-2.5"}. This record does not assert independent reproduction.

### Narrow Secret Loyalty (Qwen-2.5-32B)

Activation Rate: 70. Result ID: 6a0177f047f34ce924e2bf6e.

[Reporting paper](<https://www.sota2.com/research/paper/6a01628bdbc836b87d82307f.md>)

[Method origin paper](<https://www.sota2.com/research/paper/6a01628bdbc836b87d82307f.md>)

Recorded evaluation configuration: {"Model":"Qwen-2.5-32B (trained)","Scale":"32B","Backbone":"Qwen-2.5"}. This record does not assert independent reproduction.

### Narrow Secret Loyalty (Qwen-2.5-7B)

Activation Rate: 69.5. Result ID: 6a0177f047f34ce924e2bf6d.

[Reporting paper](<https://www.sota2.com/research/paper/6a01628bdbc836b87d82307f.md>)

[Method origin paper](<https://www.sota2.com/research/paper/6a01628bdbc836b87d82307f.md>)

Recorded evaluation configuration: {"Model":"Qwen-2.5-7B (trained)","Scale":"7B","Backbone":"Qwen-2.5"}. This record does not assert independent reproduction.

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