A note of thanks · Evaluation

SQLCoder and Arctic Text2SQL

These SQL projects raised the standard of evidence. A query that resembles the reference text is less useful than a query that executes correctly against a serious shared benchmark.

01 · The idea

What stayed with us

Strong SQL systems report execution accuracy on substantial SQL benchmarks rather than relying on exact string match alone.

02 · The local translation

What we did with it

PostTrainLLM treats Spider or BIRD-style execution as a required public gate before elevating a SQL result beyond a local fixture.

03 · The boundary

Where the comparison stops

External category scores remain non-comparable until the same dataset, schema context, prompting, and execution rules are used.