# Detoxification on REALTOXICITYPROMPTS (test)

Source: [Detoxification on REALTOXICITYPROMPTS (test)](<https://www.sota2.com/research/sota/detoxification-on-realtoxicityprompts-test>)

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

Data updated: 2026-02-26T06:07:32.002Z

Evaluation of model detoxification capabilities on the test split of the RealToxicityPrompts dataset, measured using toxicity scores, perplexity (PPL), and distribution metrics (dist-2, dist-3).

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 | Detoxification |
| Dataset (including any stated split) | REALTOXICITYPROMPTS (test) |
| Metric | Toxicity Score (Avg) |
| Unit | SCALAR |
| 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

- [Toxicity Score (Avg)](<https://www.sota2.com/research/sota/detoxification-on-realtoxicityprompts-test.md?metric=Toxicity%20Score%20(Avg)>): Lower is better; unit: SCALAR.

- [PPL](<https://www.sota2.com/research/sota/detoxification-on-realtoxicityprompts-test.md?metric=PPL>): Lower is better; unit: SCALAR.

- [Dist-2](<https://www.sota2.com/research/sota/detoxification-on-realtoxicityprompts-test.md?metric=Dist-2>): Higher is better; unit: PERCENT.

- [Dist-3](<https://www.sota2.com/research/sota/detoxification-on-realtoxicityprompts-test.md?metric=Dist-3>): Higher is better; unit: PERCENT.

## Recorded results

| Method | Toxicity Score (Avg) | Result date | Report paper |
| --- | --- | --- | --- |
| FINE-GRAINED RLHF | 0.081 | Not recorded | 68fa71021e209131cea4e83c |
| Holistic RLHF | 0.13 | Not recorded | 68fa71021e209131cea4e83c |
| DEXPERTS | 0.136 | Not recorded | 68fa71021e209131cea4e83c |
| GeDi | 0.154 | Not recorded | 68fa71021e209131cea4e83c |
| GPT-2 | 0.192 | Not recorded | 68fa71021e209131cea4e83c |

### FINE-GRAINED RLHF

Toxicity Score (Avg): 0.081. Result ID: 699187f03a6c31484908c432.

[Reporting paper](<https://www.sota2.com/research/paper/68fa71021e209131cea4e83c.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa71021e209131cea4e83c.md>)

[Implementation](<https://finegrainedrlhf.github.io/>)

Recorded evaluation configuration: {"Reward Type":"Sentence-level (fine-grained)"}. This record does not assert independent reproduction.

### Holistic RLHF

Toxicity Score (Avg): 0.13. Result ID: 699187f03a6c31484908c431.

[Reporting paper](<https://www.sota2.com/research/paper/68fa71021e209131cea4e83c.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa66e31e209131cea12d84.md>)

Recorded evaluation configuration: {"Reward Type":"Holistic"}. This record does not assert independent reproduction.

### DEXPERTS

Toxicity Score (Avg): 0.136. Result ID: 699187f03a6c31484908c430.

[Reporting paper](<https://www.sota2.com/research/paper/68fa71021e209131cea4e83c.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa6a0b1e209131cea25b2e.md>)

[Implementation](<https://github.com/alisawuffles/DExperts>)

Recorded evaluation configuration: {"Approach":"Controlled Generation"}. This record does not assert independent reproduction.

### GeDi

Toxicity Score (Avg): 0.154. Result ID: 699187f03a6c31484908c42f.

[Reporting paper](<https://www.sota2.com/research/paper/68fa71021e209131cea4e83c.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa68a21e209131cea1d517.md>)

Recorded evaluation configuration: {"Approach":"Controlled Generation"}. This record does not assert independent reproduction.

### GPT-2

Toxicity Score (Avg): 0.192. Result ID: 699187f03a6c31484908c42e.

[Reporting paper](<https://www.sota2.com/research/paper/68fa71021e209131cea4e83c.md>)

Recorded evaluation configuration: {"Model Size":"Large"}. This record does not assert independent reproduction.

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