Published June 27, 2026 | Version v0.1.0

nl2sql-onprem-bench: an honest, reproducible benchmark of on-prem open-weight LLMs for Text-to-SQL

Authors/Creators

Description

First public release. On-prem Text-to-SQL benchmark on BIRD dev (n=1534) for Qwen2.5-Coder 7B/14B/32B: base / self-correct / self-consistency / lexical schema-linking. Includes per-example predictions, per-stage summaries, and size + cost/latency frontier analyses.

Notes

If you use this benchmark or code, please cite it.

Files

beskvladimir-create/nl2sql-onprem-bench-v0.1.0.zip

Files (1.6 MB)

Additional details