OraShift
Oracle SQL to PostgreSQL with a QLoRA fine-tune of Qwen2.5-Coder-3B, and every translation judged by executing it on both databases. A translation counts only if it runs on PostgreSQL and returns what the original returns on Oracle.
Execution accuracy over 330 held-out test units. Hybrid keeps sqlglot's output when it verifies against Oracle and uses the fine-tuned model otherwise.
Paste Oracle SQL
The rules run live in your browser. This free demo does not run the 3B model; for test-set statements it shows the model's real saved output.
All 330 held-out units with the base and fine-tuned models' actual outputs. Pass or fail was decided by execution, so text that differs from the reference can still be correct. Per-construct accuracy is shown for the selected construct.
By generalisation axis
Two things are held out: the logistics schema, and 25 of
123 templates. The fine-tune generalises to unseen table names, not to unseen constructs.
By construct
Limitations
- On an unseen construct variant the fine-tune is no better than the base model, and on
SYSDATEit is far worse (2/26 against 23/26) because it echoes unfamiliar input unchanged. - The data is synthetic and template-generated, so it is more uniform than real-world SQL.
- This page cannot execute anything. Every accuracy figure comes from the offline evaluation.