# Language Model Alignment on Safe RLHF

Source: [Language Model Alignment on Safe RLHF](<https://www.sota2.com/research/sota/language-model-alignment-on-safe-rlhf>)

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

Data updated: 2026-03-26T13:28:20.100Z

Evaluation of Language Model Alignment performance on the Safe RLHF dataset, focusing on Helpfulness and Harmlessness Win Rates, Reward, and Cost 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 | Language Model Alignment |
| Dataset (including any stated split) | Safe RLHF |
| Metric | Win Rate (Helpfulness) |
| Unit | PERCENT |
| Direction | Higher is better |
| Results loaded | 3 of 3 |
| 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

- [Win Rate (Helpfulness)](<https://www.sota2.com/research/sota/language-model-alignment-on-safe-rlhf.md?metric=Win%20Rate%20(Helpfulness)>): Higher is better; unit: PERCENT.

- [Win Rate (Harmlessness)](<https://www.sota2.com/research/sota/language-model-alignment-on-safe-rlhf.md?metric=Win%20Rate%20(Harmlessness)>): Higher is better; unit: PERCENT.

- [Reward](<https://www.sota2.com/research/sota/language-model-alignment-on-safe-rlhf.md?metric=Reward>): Higher is better; unit: SCALAR.

- [Cost](<https://www.sota2.com/research/sota/language-model-alignment-on-safe-rlhf.md?metric=Cost>): Lower is better; unit: SCALAR.

## Recorded results

| Method | Win Rate (Helpfulness) | Result date | Report paper |
| --- | --- | --- | --- |
| PbCRL | 80.7 | Not recorded | 69c4bd640636e5bf4afc7435 |
| Safe RLHF | 79.4 | Not recorded | 69c4bd640636e5bf4afc7435 |
| PPO | 72.5 | Not recorded | 69c4bd640636e5bf4afc7435 |

### PbCRL

Win Rate (Helpfulness): 80.7. Result ID: 69c5345c47f34ce924aec385.

[Reporting paper](<https://www.sota2.com/research/paper/69c4bd640636e5bf4afc7435.md>)

[Method origin paper](<https://www.sota2.com/research/paper/69c4bd640636e5bf4afc7435.md>)

Recorded evaluation configuration: {"Base Model":"Llama-3.2-1B","Judge":"Gemini 2.5 Flash"}. This record does not assert independent reproduction.

### Safe RLHF

Win Rate (Helpfulness): 79.4. Result ID: 69c5345c47f34ce924aec386.

[Reporting paper](<https://www.sota2.com/research/paper/69c4bd640636e5bf4afc7435.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa70e81e209131cea4ddd5.md>)

Recorded evaluation configuration: {"Base Model":"Llama-3.2-1B","Judge":"Gemini 2.5 Flash"}. This record does not assert independent reproduction.

### PPO

Win Rate (Helpfulness): 72.5. Result ID: 69c5345c47f34ce924aec384.

[Reporting paper](<https://www.sota2.com/research/paper/69c4bd640636e5bf4afc7435.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa63a01e209131ce9ffdaa.md>)

Recorded evaluation configuration: {"Base Model":"Llama-3.2-1B","Judge":"Gemini 2.5 Flash"}. This record does not assert independent reproduction.

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