Datasets:

scottfitz
/

text-to-sql-recipe-data

---
license: cc-by-sa-4.0
pretty_name: Text-to-SQL Recipe Data
language:
- en
task_categories:
- table-question-answering
- text-generation
tags:
- text-to-sql
- sql
- bird
- reinforcement-learning
size_categories:
- 1K<n<10K
---
# Text-to-SQL Recipe Data

The datasets the River text-to-SQL reinforcement-learning recipe trains and evaluates on, laid out the way its
`setup data` command reads them. The recipe checks every file against SHA-256 pins in its code, so use the
command rather than copying files by hand. From the recipe's directory (the `text-to-sql` example of River Recipes):

```bash
pip install '.[data]'
eval "$(python -m text_to_sql_recipe setup data)"
```

## Contents

**`revisql-verified/`: BIRD-Platinum training data.**
- `dataset/`: the ReViSQL authors' verified release of BIRD training tasks, with 2,064 training and 398 validation
  tasks. Each row carries the question, evidence, gold SQL, grading method and the task text the model sees (the
  database schema, the evidence and the question).
- `hints/`: one training hint per training task, with a binding to the exact prompt, gold query and database it was
  written for. The hints were written by Kimi K2.6 at temperature 0 from each task's question, evidence, schema and
  gold query.
- `verifier/gate-policy.json`: per database, the schema and the audited foreign-key status that the optional VeriEQL
  counterexample gate reads.
- `subsets/bird-platinum-hard.json`: the 689 BIRD-Platinum-Hard training tasks, the questions strong models most often
  failed on every attempt early in reinforcement-learning training (see `scottfitz/BIRD-Platinum-Hard`).
- `databases/`: BIRD's 69 training databases, unchanged.

**`arcwise/`: Arcwise-Plat-SQL evaluation data.**
- 498 evaluation tasks over the 11 BIRD Mini-Dev databases (`databases/`).
- The gold SQL is the edition the ReViSQL paper evaluated (`uiuc-kang-lab/text_to_sql_benchmarks` at commit
  `bef1f2e42b`).
- Questions 518 and 701 use result-preserving rewrites of their gold queries, because the canonical queries exceed
  the grading time limit.

**`release.json`:** every file with its size and SHA-256.

## Sources and licences

This release is distributed under CC BY-SA 4.0, the licence of the data it derives from.

- **BIRD** (Li et al., NeurIPS 2023): the databases, questions and evidence. CC BY-SA 4.0.
- **BIRD-Platinum** (`uiuc-kang-lab/bird-platinum`): corrected BIRD training annotations. CC BY-SA 4.0. The verified
  task release used here is ReViSQL's (Jin et al., arXiv 2603.20004).
- **Arcwise-Plat-SQL:** Jin, Choi, Zhu and Kang, "Pervasive Annotation Errors Break Text-to-SQL Benchmarks and
  Leaderboards" (VLDB 2026), `uiuc-kang-lab/text_to_sql_benchmarks`. CC BY-SA 4.0. It builds on Arcwise's corrections
  of BIRD Mini-Dev.
- **Training hints and the gate policy:** ours, under CC BY-SA 4.0.

	---
	license: cc-by-sa-4.0
	pretty_name: Text-to-SQL Recipe Data
	language:
	- en
	task_categories:
	- table-question-answering
	- text-generation
	tags:
	- text-to-sql
	- sql
	- bird
	- reinforcement-learning
	size_categories:
	- 1K<n<10K
	---

	# Text-to-SQL Recipe Data

	The datasets the River text-to-SQL reinforcement-learning recipe trains and evaluates on, laid out the way its
	`setup data` command reads them. The recipe checks every file against SHA-256 pins in its code, so use the
	command rather than copying files by hand. From the recipe's directory (the `text-to-sql` example of River Recipes):

	```bash
	pip install '.[data]'
	eval "$(python -m text_to_sql_recipe setup data)"
	```

	## Contents

	`revisql-verified/`: BIRD-Platinum training data.
	- `dataset/`: the ReViSQL authors' verified release of BIRD training tasks, with 2,064 training and 398 validation
	tasks. Each row carries the question, evidence, gold SQL, grading method and the task text the model sees (the
	database schema, the evidence and the question).
	- `hints/`: one training hint per training task, with a binding to the exact prompt, gold query and database it was
	written for. The hints were written by Kimi K2.6 at temperature 0 from each task's question, evidence, schema and
	gold query.
	- `verifier/gate-policy.json`: per database, the schema and the audited foreign-key status that the optional VeriEQL
	counterexample gate reads.
	- `subsets/bird-platinum-hard.json`: the 689 BIRD-Platinum-Hard training tasks, the questions strong models most often
	failed on every attempt early in reinforcement-learning training (see `scottfitz/BIRD-Platinum-Hard`).
	- `databases/`: BIRD's 69 training databases, unchanged.

	`arcwise/`: Arcwise-Plat-SQL evaluation data.
	- 498 evaluation tasks over the 11 BIRD Mini-Dev databases (`databases/`).
	- The gold SQL is the edition the ReViSQL paper evaluated (`uiuc-kang-lab/text_to_sql_benchmarks` at commit
	`bef1f2e42b`).
	- Questions 518 and 701 use result-preserving rewrites of their gold queries, because the canonical queries exceed
	the grading time limit.

	`release.json`: every file with its size and SHA-256.

	## Sources and licences

	This release is distributed under CC BY-SA 4.0, the licence of the data it derives from.

	- BIRD (Li et al., NeurIPS 2023): the databases, questions and evidence. CC BY-SA 4.0.
	- BIRD-Platinum (`uiuc-kang-lab/bird-platinum`): corrected BIRD training annotations. CC BY-SA 4.0. The verified
	task release used here is ReViSQL's (Jin et al., arXiv 2603.20004).
	- Arcwise-Plat-SQL: Jin, Choi, Zhu and Kang, "Pervasive Annotation Errors Break Text-to-SQL Benchmarks and
	Leaderboards" (VLDB 2026), `uiuc-kang-lab/text_to_sql_benchmarks`. CC BY-SA 4.0. It builds on Arcwise's corrections
	of BIRD Mini-Dev.
	- Training hints and the gate policy: ours, under CC BY-SA 4.0.