Aurora Series v2
Insurance lead classification model trained with GRPO on Chinese customer service dialogues.
- Base: Aurora-Series-v1 (grpo_v4_merged)
- Training: 200 steps GRPO v2, peft merged (checkpoint-200, best checkpoint)
- Eval: T1=44.1%, T2=92.2%, T3=85.7%, avg=74.0% ✅
Aurora Series
| Version |
Base |
Steps |
T1 |
T2 |
T3 |
Avg |
| v1 |
sft_v5 |
500 GRPO |
— |
— |
— |
72.9% |
| v2 |
Aurora-v1 |
200 GRPO |
44.1% |
92.2% |
85.7% |
74.0% ✅ best |
