Text-to-SQL on Aggregated Spider/BIRD Splits
70.2AccuracyRefGRPO
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| RefGRPOBackbone=Qwen2.5-Coder-3B-Instruct2026.06 | 70.2 | 79.7 | 23.1 | 1.3 | 71.1 | |
| GRPO+Backbone=Qwen2.5-Coder-3B-Instruct2026.06 | 69.4 | 76.6 | 24.6 | 23.7 | 68.5 | |
| RefGRPOBackbone=Llama-3.2-3B-Instruct2026.06 | 64 | 76 | 26.3 | 23.5 | 62.2 | |
| GRPO+Backbone=Llama-3.2-3B-Instruct2026.06 | 62.9 | 74.5 | 26.5 | 30.9 | 58.4 | |
| BaseBackbone=Qwen2.5-Coder-3B-Instruct2026.06 | 56.8 | 71.4 | 31.4 | 25.2 | 53.8 | |
| BaseBackbone=Llama-3.2-3B-Instruct2026.06 | 52.1 | 65.1 | 37.5 | 31.4 | 49 |