Search Relevance Evaluation on Taobao AI Search Head Queries
61.5HRLEAPS-HR
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| LEAPS-HRBase Model=Qwen3-14B, Optimization Objective=Hybrid Relevance (HR), Training Strategy=RL2026.01 | 61.5 | 72.5 | 63.8 | |
| LEAPS-ERBase Model=Qwen3-14B, Optimization Objective=Effective Relevance (ER), Training Strategy=RL2026.01 | 59.8 | 78.5 | 68.5 | |
| LEAPS-GRBase Model=Qwen3-14B, Optimization Objective=Global Relevance (GR), Training Strategy=RL2026.01 | 57.5 | 80.5 | 66.5 | |
| SFT (Qwen3-14B)Base Model=Qwen3-14B, Training Strategy=SFT2026.01 | 48.5 | 63.8 | 55.6 | |
| SFT (Tbstar-16A2B)Base Model=Tbstar-16A2B, Training Strategy=SFT2026.01 | 48.2 | 63.4 | 55.3 | |
| Production Baseline2026.01 | 46.2 | 61.2 | 53.1 |