43K schema-linking training rows: {query, pos, neg, pos_scores, neg_scores, prompt, type} — positives are gold columns, negatives are hard negatives from the same database.
spider/train.jsonl 8,304 rows Spider train (official)
bird/train.jsonl 9,351 rows BIRD train (official)
synthetic/train.jsonl 25,324 rows SQL-GEN-style synthetic over public BigQuery/Snowflake schemas
seed_databases/ seed schemas +… See the full description on the dataset page:
https://huggingface.co/datasets/griffith-bigdata/GRAST-SL-training-set.
