Multi-Turn Consistency on Role-playing dataset
2.65KE ScorePPR-GDE
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| PPR-GDEBackbone=Qwen2.5-3B-Instruct2026.05 | 2.65 | 2.94 | 3.03 | 2.87 | |
| GRPOBackbone=Qwen2.5-3B-Instruct2026.05 | 2.61 | 3.05 | 2.89 | 2.85 | |
| PPR-GDEBackbone=Qwen2.5-1.5B-Instruct2026.05 | 2.57 | 3.01 | 2.99 | 2.86 | |
| PPOBackbone=Qwen2.5-3B-Instruct2026.05 | 2.48 | 3.1 | 2.93 | 2.84 | |
| GRPOBackbone=Qwen3-0.6B2026.05 | 2.45 | 2.86 | 2.64 | 2.65 | |
| PPOBackbone=Qwen3-0.6B2026.05 | 2.42 | 2.89 | 2.68 | 2.66 | |
| PPR-GDEBackbone=Qwen3-0.6B2026.05 | 2.39 | 3.06 | 3.04 | 2.83 | |
| PPOBackbone=Qwen2.5-1.5B-Instruct2026.05 | 2.37 | 2.93 | 2.68 | 2.66 | |
| GRPOBackbone=Qwen2.5-1.5B-Instruct2026.05 | 2.33 | 2.93 | 2.85 | 2.7 | |
| w/o RLBackbone=Qwen3-0.6B2026.05 | 2.23 | 2.68 | 2.61 | 2.51 | |
| w/o RLBackbone=Qwen2.5-3B-Instruct2026.05 | 2.17 | 2.83 | 2.95 | 2.65 | |
| w/o RLBackbone=Qwen2.5-1.5B-Instruct2026.05 | 1.82 | 2.7 | 2.79 | 2.44 |