antgroup/Agentar-Scale-SQL

Implements orchestrated test-time scaling with three synergistic perspectives: a Light Schema Engine for efficient database representation, a vector database pipeline (Chroma) for semantic retrieval of training examples and cell values, and BM25 indexing for content-based schema matching. The framework includes SQL candidate generation via ICL and reasoning generators, iterative refinement, and a selection module that ranks outputs—achieving 81.67% execution accuracy on the BIRD benchmark. Provides pre-trained 32B generation models on Hugging Face/ModelScope, along with open-source preprocessing pipelines and schema engines for custom database adaptation.

372 stars.