metadata
license: cc-by-sa-4.0
CoRE-DB evaluation sets (RQ3, RQ6)
Evaluation sets of Small Language Models for Text-to-SQL against Spurious Correctness.
| folder | paper use | content |
|---|---|---|
rq3/spurious_train200_3b/ |
RQ3 component table (3B) | 100 BIRD + 100 Spider train questions, 800 RL rollouts of the 3B CoRE-DB model, 3-LLM majority labels |
rq6/rollout_sample300/ |
RQ6 unseen-operation panel | 300 sampled training rollouts with unseen atomic operations, blind 3-LLM labels |
rq6/direct_neighbor_sources/ |
RQ6 coverage panel | grammar and LLM neighbour queries per DB, coverage summary |
CASES.jsonl / CASE_RESULTS.jsonl hold one record per case; MANIFEST.json and JUDGING_PROTOCOL.json describe sampling
and labelling. Databases: thanhdath/core-db (CoRE-DB) and the official BIRD / Spider train DBs.
RQ6 public inputs (2026-11-01)
Used by scripts/rq6/ of the CoRE-DB code release to re-derive the RQ6 subset and both RQ6 tables without any LLM call.
| File | Content |
|---|---|
rq6/neighbor_pool/pairs-eligible.parquet |
252,036 gold-neighbor pairs over the 215 training databases (grammar mutations 156,846 + LLM-written SQL 95,190) that match on the original database (label tied, provenance ok, no execution flag); columns dataset, db_id, generator, gold_sql, neighbor_sql, label, provenance_check, exec_flags |
rq6/coverage_rollouts/core12_05b_rollouts.jsonl.gz |
96,000 rollouts (3,000 steps x 32, steps 0-2999) of the 0.5B CoRE-DB training run; fields ds, db, step, sample_id, sql |
The RQ6 evaluation subset itself (210 of the 800 RQ3 cases) ships with the code as
evaluation_sets/rq6_unseen210_from_3b.jsonl. rq6/rollout_sample300/ is an earlier sample that the paper no longer uses.