Run archive for "What Does Execution-Gated Refinement Buy in Text-to-SQL? Cost, Accuracy Across Backbones, and Negative Results"
Abstract
Stored requests and responses, predictions and per-run metadata of the experiments reported in "What Does Execution-Gated Refinement Buy in Text-to-SQL? Cost, Accuracy Across Backbones, and Negative Results" (Li, Song, Wang; submitted to Applied Sciences). 161 runs, 1,401 files (1.2 GB unpacked). The archive allows the analyses and tables of the paper to be recomputed without new model calls; the code and the replay instructions are in the repository https://github.com/vivia-an/nl2sql-gated-refinement (see its README). File integrity: SHA256SUMS. Licence and modification notice: the archive contains questions, evidence and schema text from BIRD, Spider and Spider-Realistic, embedded in model requests and responses; this is a modification of that material, shared under CC BY-SA 4.0 with attribution to the original authors. The underlying BIRD databases are not redistributed and keep their upstream licences. Model responses (Qwen models via Alibaba Cloud Model Studio / DashScope, paid account) are shared as research data with attribution to the provider; no additional licence is claimed on the raw output text, and its use remains subject to the provider's terms.
