thanhdath's picture
Keep only the evaluation set: one file, 800 gold-prediction pairs
1dc8831 verified
|
Raw History Blame Contribute Delete
930 Bytes
---
license: cc-by-sa-4.0
configs:
- config_name: default
data_files:
- split: test
path: evaluation_set.jsonl
---
# CoRE-DB spurious-correctness evaluation set
Spurious-correctness evaluation set of *Boosting Small Language Models for Text-to-SQL with Cost-Efficient Reinforcement Environment against Spurious Correctness*: 800 pairs of a gold SQL and a predicted SQL (`gold_sql`, `pred_sql`), four GRPO rollouts of FINER-SQL 3B for each of 100 BIRD and 100 Spider training questions (`ds`, `db`, `question`, `evidence`), over 107 databases.
`label` is `spurious` (152 pairs: same result as the gold SQL on the original database but not semantically equivalent), `correct` (325) or `wrong` (323), where `correct` and `wrong` are the 648 non-spurious pairs; `unseen` marks the 210 predictions with an atomic operation absent from the neighbor queries used to construct CoRE-DB.
Code: https://github.com/thanhdath/core-db