Text-to-SQL on AmbiQT in-domain
92.3Single Interpretation CoverageOurs
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| OursApproach=Disambiguate-and-Parse, Method=Interpretation Infilling, Backbone=Llama-3.1 8B2025.02 | 92.3 | 53.2 | 72.8 | 38.2 | |
| Gold Interp. SFTApproach=Disambiguate-and-Parse, Training=SFT on interpretations, Backbone=Llama-3.1 8B2025.02 | 87.4 | 61.2 | 74.1 | 79.6 | |
| SFTApproach=End-to-End Text-to-SQL, Training=SFT with LoRA, Backbone=Llama-3.1 8B2025.02 | 82.1 | 63.2 | 72.7 | 74.3 | |
| Interp. PromptApproach=Disambiguate-and-Parse, Method=LLM Prompting, Backbone=Llama-3.1 8B2025.02 | 81.8 | 26 | 53.4 | 31.8 | |
| w. Self-CorrectionApproach=Disambiguate-and-Parse, Self-Correction=true, Backbone=Llama-3.1 8B2025.02 | 77.4 | 13.9 | 45.7 | 46 | |
| 0-shot PromptApproach=End-to-End Text-to-SQL, Shots=0, Backbone=Llama-3.1 8B2025.02 | 62.3 | 12.3 | 37.3 | 58.1 | |
| 3-shot PromptApproach=End-to-End Text-to-SQL, Shots=3, Backbone=Llama-3.1 8B2025.02 | 44.3 | 10.9 | 27.6 | 33 |