BIRD-Platinum-Hard

BIRD-Platinum-Hard is a 689-question subset of the BIRD-Platinum training questions, across 67 databases. It keeps the questions that strong models most often failed to solve on every attempt early in reinforcement-learning training. A pass over it samples about a third as many questions as an epoch over the full 2,064-question training split.

How the questions were selected

We trained four models with our text-to-SQL recipe: GLM-5.3 Flash, Kimi K2.6, DeepSeek V4.1 Flash and Qwen3.8-27B. Each training question was attempted 16 times on its first visit. A question is in BIRD-Platinum-Hard if fewer than three of the four models answered all 16 attempts correctly.

Models that solved it 16/16	Questions	Included
0	362	yes
1	127	yes
2	200	yes
3	368	no
4	1,007	no

One first visit is a noisy measurement: two runs of the same model agree on about 87–91% of the questions they solve 16/16. Read the set as "often not fully solved", not "never solved".

In our experiments, training on this subset improved faster early on and plateaued sooner than training on the full split.

Fields

The records are exactly the rows our recipe trained on: the ReViSQL authors' verified release of BIRD-Platinum (bird-verified-train.json in the ReViSQL repository). That release corrects some records of the candidate file published as uiuc-kang-lab/bird-platinum. Within this subset, 11 questions, 17 evidence texts and 15 gold queries differ from the candidate file, and 627 questions carry a different grading method (mostly set where the candidate file says multiset). Question 1104 differs from the authors' file itself, in its grading annotation only; we publish the value we trained with.

• question_id: the BIRD question id, also BIRD-Platinum's id
• db_id: the BIRD training database the question runs against
• question, evidence, SQL: the question, its external knowledge and the gold query
• grading_method: how query results are compared (for example set, or subset,=,5)

The databases are BIRD's training databases, available from the BIRD benchmark. They are not included here.

The split config

The BIRD-Platinum candidate file has 2,463 records and no partition. Our recipe uses the train/validation partition from the ReViSQL authors' verified release: 2,064 training and 398 validation questions (1 upstream record is in neither). The split config lists each question_id with its side. Every BIRD-Platinum-Hard question is on the training side.

Usage

from datasets import load_dataset

hard = load_dataset("scottfitz/BIRD-Platinum-Hard", split="train")
partition = load_dataset("scottfitz/BIRD-Platinum-Hard", "split", split="train")

License and attribution

CC BY-SA 4.0. This dataset is a subset of the ReViSQL authors' verified release of BIRD-Platinum (Zhu et al., 2026), which is derived from BIRD (Li et al., 2023). BIRD and BIRD-Platinum are licensed CC BY-SA 4.0. Changes: we selected a subset of questions; the records are as we trained on them, described under Fields.

Citation

@article{zhu2026revisql,
  title   = {Human-Level Text-to-SQL via Reinforcement Learning on Verified Data, Without Pipeline Engineering},
  author  = {Zhu, Yuxuan and Jin, Tengjun and Choi, Yoojin and Kang, Daniel},
  journal = {arXiv preprint arXiv:2603.20004},
  year    = {2026}
}

@inproceedings{li2023bird,
  title     = {Can LLM Already Serve as A Database Interface? A BIg Bench for Large-Scale Database Grounded Text-to-SQLs},
  author    = {Li, Jinyang and others},
  booktitle = {Advances in Neural Information Processing Systems},
  year      = {2023}
}