SREB: SQL Rewrite Evidence Benchmark

A dataset of executable evidence for SQL query rewrites proposed by language models.

Most SQL benchmarks ask "write a query for this question". Most rewrite papers report a success rate. SREB publishes something different: for every rewrite, the database state on which the result changes and the rows that prove it.

reference [1,1,1,1]    candidate [1,1,1]    ->  multiplicities lost

Why it is different

Existing work	What it gives	Missing
Spider, BIRD	text-to-SQL evaluation	no rewrite validation
GenRewrite, R-Bot, LLM-R2	success rates on benchmarks	no per-rewrite witness
Cosette, VeriEQL	formal proofs for supported SQL	coverage, and no run-level evidence

SREB keeps the evidence, not the score. Every record is reproducible: database version, model digest, sampling parameters, state profile, and the rows that differ.

What is inside

Config	Content
corpus	query families: schema DDL, constraints, reference SQL, breaking profile
states	frozen state banks generated under the real schema constraints
runs	model responses with digest, quantization and sampling parameters
evidence	outcome per (family, model): status, refuting states, row witness

Outcome values: refuted, preserved, no_rewrite, unsupported, error.

preserved means no refuting state was found on the finite bank. It is not a proof of equivalence, and the dataset never claims otherwise.

Reproduce

git clone https://github.com/ykvlev/sql-rewrite-evidence-bench
pip install -r scripts/requirements.txt
python scripts/check_corpus.py --repo-root .        # hashes, pins, integrity
python scripts/recompute_tables.py --repo-root .    # rebuild summary tables

Every published number is recomputable from data/ without rerunning a model.

Scale (v0.3.0)

• query families: 204 across three schemas (library, online shop, university)
• frozen states: 5 712 (28 profiles per family)
• model runs: 1 224 (6 models × 204 evaluated families)
• evidence: 204 families, evaluated on MySQL 8.4.11
• total checks: 34 272 (model × family × state)
• models: Qwen2.5-Coder-7B, Qwen2.5-7B, Qwen2.5-3B, Phi-3.5-mini-3.8B, DeepSeek-Coder-V2, Llama-3.2-3B

Cross-DBMS check: the schemas and the 136 library/shop reference queries execute on PostgreSQL 16.15 without modification; on the 12 seed families the reference results and candidate classifications match MySQL exactly (0 divergences over 336 state pairs). The university schema has not been ported yet.

Limitations

• Synthetic schemas: coverage of real workloads is partial.
• A finite state bank cannot prove equivalence.
• Witnesses are not minimised: the recorded state is the first one that refuted the candidate, not the smallest possible.
• Model behaviour changes across engine versions; the engine version is part of the record. The six models span Ollama 0.40.4–0.40.6.
• Small models can return invalid SQL: 23 answers are unsupported and 33 are error of 1 224 (recorded, not retried).

Licence

Data and documentation: CC BY 4.0. Code: MIT.

Citation

@dataset{yakovlev_sreb,
  title  = {SREB: SQL Rewrite Evidence Benchmark},
  author = {Yakovlev, Artem},
  year   = {2026},
  note   = {ORCID 0009-0006-8617-7895},
  url    = {https://github.com/ykvlev/sql-rewrite-evidence-bench}
}