devansh173/orashift-pairs

OraShift translation pairs

Oracle SQL to PostgreSQL translation pairs, every one verified by execution: the Oracle statement was run on Oracle, the PostgreSQL statement on PostgreSQL, and the results compared. A pair is only present if they matched.

Measured against Oracle AI Database 26ai Free 23.26.2.0.0 and PostgreSQL 17.10. Both the data and the schemas are synthetic.

Splits

Split	Examples	What it measures
train	418	Training. Seen schemas, seen templates.
val	35	Validation, carved from the same bucket as training.
test_in_schema	33	Test. Seen schema, seen template: the easiest case.
test_unseen_schema	160	Test. Held-out logistics schema, template seen in training.
test_unseen_template	102	Test. Seen schema, construct variant never trained on.
test_unseen_both	35	Test. Held-out schema and held-out template: the hardest case.
total	783

Why there are four test sets

Holding out one schema measures less than it appears to. If a template is seen during training as retail and tested as logistics, the model has already met that exact translation pattern and only the table names are new -- that tests vocabulary robustness, not translation skill.

So two things are withheld: the whole logistics schema, and 25 of 123 templates. Reporting the resulting four cells separately shows which kind of generalisation actually holds.

Format

Chat JSONL, one example per line:

{"messages": [
  {"role": "system", "content": "task rules + the Oracle DDL this statement uses"},
  {"role": "user", "content": "<Oracle statement>"},
  {"role": "assistant", "content": "<PostgreSQL statement>"}
], "meta": {"schema": "...", "category": "...", "unit_type": "..."}}

The system prompt carries only the table definitions the statement actually references, not the whole schema: the rest would be spent token budget. The assistant turn is the translation alone, with no explanation, so loss on assistant tokens trains exactly the output wanted at inference.

Coverage

Category	train	val	test_in_schema	test_unseen_schema	test_unseen_template	test_unseen_both
add_months	40	5	4	17	1	1
aggregate	24	3	2	10	2	1
analytic	16	1	1	6	3	1
connect_by	8	1	1	5	2	1
cte	6	0	0	2	3	1
ddl_constraint	4	0	0	1	1	0
ddl_identity	2	0	0	0	0	0
ddl_index	4	0	0	1	0	0
ddl_sequence	3	0	0	0	1	0
ddl_table	4	0	0	1	0	0
ddl_view	5	0	0	2	3	1
decode	8	0	0	4	2	0
dml_delete	3	0	0	1	3	1
dml_insert	3	0	0	1	0	0
dml_update	7	0	0	3	0	0
dual	11	1	1	0	1	0
fetch_first	64	7	7	26	12	4
listagg	6	0	0	2	2	1
merge	3	0	0	1	3	1
minus	9	0	0	3	3	1
null_concat	8	0	0	3	2	1
nvl	16	1	1	7	1	1
outer_join_plus	9	0	0	3	3	1
pivot	2	0	0	1	3	1
rownum	51	6	6	21	3	1
sequence	6	0	0	2	0	0
string_func	24	3	3	10	3	1
subquery	10	1	1	4	3	1
sysdate	7	0	0	2	18	6
to_char_date	34	4	4	12	21	7
trunc_date	21	2	2	9	3	1

Honest limitations

Thin categories. These have fewer than 10 training examples, so any per-category result for them is noisy and should not be read as a trend:

• pivot: 2 training examples
• ddl_identity: 2 training examples
• ddl_sequence: 3 training examples
• dml_delete: 3 training examples
• dml_insert: 3 training examples
• merge: 3 training examples
• ddl_constraint: 4 training examples
• ddl_index: 4 training examples
• ddl_table: 4 training examples
• ddl_view: 5 training examples
• cte: 6 training examples
• listagg: 6 training examples
• sequence: 6 training examples
• dml_update: 7 training examples
• sysdate: 7 training examples
• connect_by: 8 training examples
• decode: 8 training examples
• null_concat: 8 training examples
• minus: 9 training examples
• outer_join_plus: 9 training examples

DDL and DML are under-represented relative to queries, because growing the pool by widening template parameters multiplies query variants far more than it multiplies schema statements.

The translations are one person's style. Each is execution-verified, so it is correct, but it is one correct style among several. A model trained on it may be more brittle to alternative phrasings than the accuracy figures suggest.

The statements are template-generated, so their phrasing is more uniform than human-written SQL.

Token lengths at build time are estimates. The build assumes 3.0 characters per token, which over-counts for SQL, so the filter errs towards excluding. The training notebook re-measures every split with the real Qwen tokenizer; those figures are in results/train_run.json (longest example 804 tokens, none excluded).

• Longest example: ~1087 estimated tokens
• Budget: 2048 tokens
• Excluded as too long: 0

Source schemas

Schema	Tables	Role
retail	5	training
hr	5	training
library	5	training
logistics	5	held out entirely

Licence

MIT. All schemas and data are synthetic; no proprietary SQL is included.