IAMIbrahim/luthorsql-bird

LuthorSQL (two-adapter System 1)

LoRA adapters and heads on seeklhy/OmniSQL-7B for LuthorSQL, a one-pass text-to-SQL model that never writes SQL token by token.

• r1/: generator. Slot heads fill a typed query plan in one forward pass; the code lists up to 32 plans and compiles them to SQLite.
• valhard_rr/: validator. Scores each executed candidate; the answer is the top-scoring candidate that runs and returns a non-null result.

Status: private pre-release. The code is in the GitHub repository IAMIbrahimmemon/luthorsql (private).

	EX
BIRD dev (original gold, 1,534)	59.39
BIRD hidden test (run by the BIRD team, 1,789)	68.03

Speed: 0.84 s per question (median, one H100 PCIe).

Training data: BIRD train, Spider train, a SynSQL-2.5M sample; the validator also used BIRD-Platinum labels, for which no license was found. License: not decided; expected research use only. Other SQL engines: without retraining, its SQLite output translated to Snowflake, BigQuery, MySQL, PostgreSQL and DuckDB scores 57.1-59.3% on BIRD dev (SQLite 60.17%). 8-bit / Apple silicon: about 9 GB with both adapters on one 8-bit backbone, same accuracy. As System 1 of a cascade with Claude Opus 5.5 through the API (threshold 3.0, forwarding only unsure questions): 33.3% cheaper than Opus alone at equal accuracy on 227 questions over databases published after Opus's training cutoff (registered test). Limitations: English, foreign-key joins only (joins use declared and name-inferred keys). The plan grammar covers 84.6% of BIRD dev gold queries.

Citation

Ibrahim Memon, Armin Karamzade, Hanif Jhaveri and Roy Fox. 2026. LuthorSQL: A One-Pass Text-to-SQL Model as System 1 for a Slower LLM Agent. Preprint. Code and data: https://github.com/IAMIbrahimmemon/luthorsql