| Dataset Name | SOTA Method | Metric | Trend | ||
|---|---|---|---|---|---|
| OpenAssistant | DPPrefSyn | GPT-4o Win Rate87.08 | 45 | 1mo ago | |
| HelpSteer (test) | PD | AlpacaEval 2 WR8.34 | 27 | 4mo ago | |
| Harmlessness | AlignX | WR87.85 | 27 | 2mo ago | |
| AlpacaEval 2.0 | GRM + LP | LC Win Rate61.52 | 25 | 2mo ago | |
| Ultrafeedback | Win Rate86 | 24 | 1mo ago | ||
| Evol-Instruct | HighAvg. | Win Rate51.4 | 24 | 3mo ago | |
| AlpacaEval | HighAvg. | Win Rate25.24 | 24 | 2mo ago | |
| HH-RLHF (test) | RE-Control | Diversity0.87 | 23 | 2mo ago | |
| HHRLHF | AAD | Average Oracle Reward-0.02 | 20 | 1mo ago | |
| EduFeedback | COALA | TFLOPs152.56 | 20 | 2mo ago | |
| UltraFeedback (test) | PD (ours) | AlpacaEval 2 Win Rate (WR)21 | 18 | 4mo ago | |
| HH-RLHF 300 prompts | CARDS | Win/Tie Rate vs Vanilla (GPT-4o)69.8 | 16 | 4mo ago | |
| SHP | RE-CONTROL + Prompting | Diversity89.3 | 15 | 4mo ago | |
| Taobao Live proprietary fine-grained preference dataset | PD (ours) | Win Score1.53 | 13 | 4mo ago | |
| AlpacaEval ULTRA 2.0 (test) | COALA | LC WR (%)31.58 | 12 | 2mo ago | |
| AlpacaEval IMDB 2.0 (test) | COALA | LC WR39.72 | 12 | 2mo ago | |
| AlpacaEval EDU 2.0 (test) | COALA | LC WR40.9 | 12 | 2mo ago | |
| AlpacaEval Length-Controlled (test) | UNA-score (MSE) | LC Win Rate8.78 | 12 | 2mo ago | |
| Gemma-3-4B | Nash Prox | Win Rate94.33 | 12 | 4mo ago | |
| PKU-SafeRLHF | AdaBoN | BWR (Median)49 | 12 | 4mo ago | |
| Alpaca, BeaverTails, and TruthfulQA (test) | AlignX | Win Rate97.1 | 12 | 4mo ago | |
| Combined Suite Setup 3 | AMA Reweighting | Average Percentage Score54.38 | 9 | 4mo ago | |
| Capybara | Win Rate28.34 | 8 | 1mo ago | ||
| Anthropic-HH (test) | DPPrefSyn | GPT-4o Win Rate57.53 | 8 | 1mo ago | |
| UltraFeedback (in-domain) | GEB-π | Win Rate (KL, alpha=1)80.6 | 8 | 4mo ago |
