Tool-Calling and Answer Generation on Almita Out-of-Domain
85.23Action RecallQwen3-1.7B + RL
Evaluation Results
| Method | Links | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| Qwen3-1.7B + RLModel Size=1.7B, Training=Reinforcement Learning (Reasoning-Action Synergy), Thinking Stage=Reasoning included2025.12 | 85.23 | 74.7 | 73.47 | 74.08 | 75.27 | 72.27 | 88.28 | 89.92 | 89.09 | 77.51 | |
| Qwen3-1.7B + SFT (Cold Start - think)Model Size=1.7B, Training=Cold-start Supervised Fine-Tuning, Thinking Stage=Reasoning included2025.12 | 84.52 | 70.57 | 76.42 | 73.38 | 68.18 | 71.18 | 90.56 | 87.65 | 89.08 | 73.15 | |
| Qwen3-1.7B + SFT (no think)Model Size=1.7B, Training=Supervised Fine-Tuning (no thinking), Thinking Stage=None2025.12 | 83.65 | 75.18 | 71.92 | 73.51 | 73.99 | 71.74 | 87.24 | 88.99 | 88.11 | 77.23 | |
| Qwen3-1.7BModel Size=1.7B, Training=Base2025.12 | 67 | 79.02 | 48.36 | 60 | 76.1 | 77.21 | 63.5 | 87.5 | 73.59 | 67.5 |