# Offline Constrained RLHF on PKU-SafeRLHF 74k (train)

Source: [Offline Constrained RLHF on PKU-SafeRLHF 74k (train)](<https://www.sota2.com/research/sota/offline-constrained-rlhf-on-pku-saferlhf-74k-train>)

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

Data updated: 2026-04-02T08:43:04.679Z

Evaluation of model performance using Offline Constrained Reinforcement Learning from Human Feedback (RLHF) methodology on the PKU-SafeRLHF 74k training set, focusing on safety, helpfulness, and violation rates.

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 | Offline Constrained RLHF |
| Dataset (including any stated split) | PKU-SafeRLHF 74k (train) |
| Metric | Safety Expected Reward (r1) |
| Unit | SCALAR |
| Direction | Higher is better |
| Results loaded | 2 of 2 |
| 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

- [Safety Expected Reward (r1)](<https://www.sota2.com/research/sota/offline-constrained-rlhf-on-pku-saferlhf-74k-train.md?metric=Safety%20Expected%20Reward%20(r1)>): Higher is better; unit: SCALAR.

- [Helpfulness Expected Reward (r2)](<https://www.sota2.com/research/sota/offline-constrained-rlhf-on-pku-saferlhf-74k-train.md?metric=Helpfulness%20Expected%20Reward%20(r2)>): Higher is better; unit: SCALAR.

- [Violation Rate](<https://www.sota2.com/research/sota/offline-constrained-rlhf-on-pku-saferlhf-74k-train.md?metric=Violation%20Rate>): Lower is better; unit: PERCENT.

## Recorded results

| Method | Safety Expected Reward (r1) | Result date | Report paper |
| --- | --- | --- | --- |
| πλ | -1.6034 | Not recorded | 69cdf7f20636e5bf4afcc99b |
| π0 | -1.8997 | Not recorded | 69cdf7f20636e5bf4afcc99b |

### πλ

Safety Expected Reward (r1): -1.6034. Result ID: 69ce2bef47f34ce924b75d86.

[Reporting paper](<https://www.sota2.com/research/paper/69cdf7f20636e5bf4afcc99b.md>)

[Method origin paper](<https://www.sota2.com/research/paper/69cdf7f20636e5bf4afcc99b.md>)

Recorded evaluation configuration: {"Policy Type":"Dual-only learned policy","Jmin":"−0.6473","η":"0.3","T":"1000"}. This record does not assert independent reproduction.

### π0

Safety Expected Reward (r1): -1.8997. Result ID: 69ce2bef47f34ce924b75d85.

[Reporting paper](<https://www.sota2.com/research/paper/69cdf7f20636e5bf4afcc99b.md>)

Recorded evaluation configuration: {"Policy Type":"Reference Policy","Jmin":"−0.6473","η":"0.3","T":"1000"}. This record does not assert independent reproduction.

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