# Preference Alignment on Koala (GPT-4o-mini & RM Evaluators)

Source: [Preference Alignment on Koala (GPT-4o-mini & RM Evaluators)](<https://www.sota2.com/research/sota/preference-alignment-on-koala-gpt-4o-mini-and-rm-evaluators>)

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

Data updated: 2026-05-11T08:29:30.220Z

Evaluation of model preference alignment performance on the Koala dataset, specifically measured using Win Rates calculated by a Reward Model and GPT-4o-mini.

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 | Preference Alignment |
| Dataset (including any stated split) | Koala |
| Metric | Win Rate (Reward Model) |
| 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 (Reward Model)](<https://www.sota2.com/research/sota/preference-alignment-on-koala-gpt-4o-mini-and-rm-evaluators.md?metric=Win%20Rate%20(Reward%20Model)>): Higher is better; unit: PERCENT.

- [Win Rate (GPT-4o-mini)](<https://www.sota2.com/research/sota/preference-alignment-on-koala-gpt-4o-mini-and-rm-evaluators.md?metric=Win%20Rate%20(GPT-4o-mini)>): Higher is better; unit: PERCENT.

- [Average Win Rate](<https://www.sota2.com/research/sota/preference-alignment-on-koala-gpt-4o-mini-and-rm-evaluators.md?metric=Average%20Win%20Rate>): Higher is better; unit: PERCENT.

## Recorded results

| Method | Win Rate (Reward Model) | Result date | Report paper |
| --- | --- | --- | --- |
| Vanilla Baseline | 77.75 | Not recorded | 6a016257dbc836b87d822eb0 |
| p-soup & Direct Fine-tuning | 70.63 | Not recorded | 6a016257dbc836b87d822eb0 |
| Direct Prompting | 52.75 | Not recorded | 6a016257dbc836b87d822eb0 |

### Vanilla Baseline

Win Rate (Reward Model): 77.75. Result ID: 6a01935347f34ce924e364d2.

[Reporting paper](<https://www.sota2.com/research/paper/6a016257dbc836b87d822eb0.md>)

Recorded evaluation configuration: {"Preference Dimensions":"2-dim"}. This record does not assert independent reproduction.

### p-soup & Direct Fine-tuning

Win Rate (Reward Model): 70.63. Result ID: 6a01935347f34ce924e364d3.

[Reporting paper](<https://www.sota2.com/research/paper/6a016257dbc836b87d822eb0.md>)

[Method origin paper](<https://www.sota2.com/research/paper/6a016257dbc836b87d822eb0.md>)

Recorded evaluation configuration: {"Preference Dimensions":"2-dim"}. This record does not assert independent reproduction.

### Direct Prompting

Win Rate (Reward Model): 52.75. Result ID: 6a01935347f34ce924e364d1.

[Reporting paper](<https://www.sota2.com/research/paper/6a016257dbc836b87d822eb0.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa70e51e209131cea4dc2d.md>)

Recorded evaluation configuration: {"Preference Dimensions":"2-dim"}. This record does not assert independent reproduction.

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