Multi-turn Dialogue on NPC-Chat (test)
3.84FluencyAT-GRPO
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| AT-GRPOEval. Type=LLM-based, Base Model=Qwen2.5-14B2026.02 | 3.84 | 3.82 | 3.41 | 3.82 | — | 3.72 | |
| GRPOEval. Type=LLM-based, Base Model=Qwen2.5-14B2026.02 | 3.83 | 3.82 | 3.37 | 3.76 | — | 3.7 | |
| SFTEval. Type=LLM-based, Base Model=Qwen2.5-14B2026.02 | 3.82 | 3.8 | 3.42 | 3.7 | — | 3.69 | |
| DeepSeekEval. Type=LLM-based, Base Model=Qwen2.5-14B2026.02 | 3.8 | 3.72 | 3.28 | 3.61 | — | 3.6 | |
| DeepSeekEval. Type=Human, Base Model=Qwen2.5-14B2026.02 | 3.1 | 3.04 | 2.84 | 2.56 | 3.52 | 3.01 | |
| GRPOEval. Type=Human, Base Model=Qwen2.5-14B2026.02 | 3 | 3.01 | 2.9 | 2.72 | 4.6 | 3.25 | |
| AT-GRPOEval. Type=Human, Base Model=Qwen2.5-14B2026.02 | 3 | 3.08 | 3.01 | 2.77 | 4.95 | 3.36 | |
| SFTEval. Type=Human, Base Model=Qwen2.5-14B2026.02 | 2.99 | 2.99 | 2.92 | 2.79 | 4.53 | 3.24 |