finer-sql
finer-sql-train-joint-fixed
finer-sql-train-joint-fixed — the FINER-SQL joint BIRD+Spider training corpus with its reference SQLs re-attached to the right sample
18,087 training rows (9,428 BIRD train + 8,659 Spider train) over 215 training databases.
Same rows, same order, same prompts as
data/train/grpo_sql_writer_origdata_joint_oldprompt_full — only the reference SQLs in
groundtruth_sqls[1:] differ.
Trainer exposure
grpo_writer.py:1264 binds the list. The binary execution match uses entry… See the full description on the dataset page: https://huggingface.co/datasets/thanhdath/finer-sql-train-joint-fixed.finer-sql-gold-neighbour-pairs
pairset/ — a million (gold, neighbour) SQL pairs over the whole BIRD+Spider training set
What this is, in one paragraph
A pair is a gold SQL query from the training corpus and a neighbour: another query
that is close to it but means something else. The point of a pair is what happens when you
run both on a database. If the two give the same answer, the execution reward that
trains a text-to-SQL model pays full marks for the wrong query — that is spurious… See the full description on the dataset page: https://huggingface.co/datasets/thanhdath/finer-sql-gold-neighbour-pairs.