Datasets:

rwugermany
/

sql-create-context-refined

license: cc-by-4.0
language:
  - en
pretty_name: sql-create-context refined
size_categories:
  - 10K<n<100K
task_categories:
  - text-generation
  - question-answering
tags:
  - text-to-sql
  - SQL
  - refinement
  - spider
source_datasets:
  - philschmid/sql-create-context-copy
  - b-mc2/sql-create-context
dataset_info:
  config_name: default
  features:
    - name: question
      dtype: string
    - name: answer
      dtype: string
    - name: context
      dtype: string
    - name: dialect
      dtype: string
    - name: spider_hardness
      dtype: string
    - name: extended_hardness
      dtype: string
    - name: classification_status
      dtype: string
    - name: syntax_status
      dtype: string
    - name: ddl_syntax_status
      dtype: string
    - name: schema_status
      dtype: string
    - name: syntax_errors
      sequence: string
    - name: ddl_syntax_errors
      sequence: string
    - name: schema_errors
      sequence: string
    - name: classification_errors
      sequence: string
    - name: status_history
      sequence: string
    - name: correction_status
      dtype: string
    - name: component1
      dtype: int64
    - name: component2
      dtype: int64
    - name: others
      dtype: int64
    - name: validation_source
      struct:
        - name: split
          dtype: string
        - name: row_index
          dtype: int64
  splits:
    - name: train
      num_bytes: 70395555
      num_examples: 78577
configs:
  - config_name: default
    data_files:
      - split: train
        path: sql_context_v2.json

sql-create-context refined

This dataset is a refinement of philschmid/sql-create-context-copy. It is not a new collection of questions. The copy is itself a duplicate of b-mc2/sql-create-context. Row order and the 78,577 questions are preserved. The SQL answers are preserved except for quote normalisation and 541 structural repairs. The CREATE TABLE context is preserved or expanded, and every row has extended_hardness. Where the official Spider adapter accepts the statement, the row also has spider_hardness.

The dataset file is data/sql/sql_context_v2.json in the sql_tools repository. On the Hub the same file is sql_context_v2.json, next to this README.md.

What changed relative to sql-create-context-copy

Step	What it does to the copy	Rows affected
Quote rewrite	Double-quoted string literals in `answer` become single-quoted SQL strings. An apostrophe inside a literal is doubled. `question` and `context` are copied unchanged. Three `FROM "t"` table names stay double-quoted.	62,967 statements changed; 503 apostrophes doubled
Validation	Generic SQLGlot check of SQL syntax, DDL syntax, and schema references. No SQL is executed.	78,036 valid; 541 rejected
Repair	Only the 541 rejected rows are rewritten. A candidate is kept only when the new SQL and DDL validate.	541 applied
DDL expansion	A model may add columns and give bare `VARCHAR` a length. The SQL query is not changed. Failed expansions keep the previous DDL.	68,983 expanded; 8,025 unchanged; 1,569 failed, all kept
Hardness	`spider_hardness` where the official adapter accepts the statement, and `extended_hardness` on every row.	78,577 extended labels; 75,382 Spider labels

dialect is generic on the 78,036 rows that passed the first validation, and mysql on the 541 repaired rows. correction_status is corrected on those 541 rows and empty otherwise. Compare spider_hardness and extended_hardness only when both are set. They use the same names (easy, medium, hard, extra). extended_hardness can also be extreme.

Hardness counts

Spider shares are of the 75,382 rows with a Spider label. Extended shares are of all 78,577 rows.

Label	Spider	Extended
easy	51,082	53,001
medium	22,693	23,957
hard	913	921
extra	694	690
extreme		8
labeled	75,382	78,577

classification_status is classified when both labels are set (75,382 rows) and extended when only extended_hardness is set (3,195 rows). No row lacks an extended label. The last entry of status_history is the same value.

Columns

Training and evaluation use question, answer, context, dialect, spider_hardness, and extended_hardness.

Column	Meaning
`question`	Natural-language question, unchanged from the copy.
`answer`	SQL query. Quotes are normalised. SQL text changes only on the 541 repaired rows.
`context`	One or more `CREATE TABLE` statements. Expanded rows have a richer DDL; unchanged and failed rows keep the previous DDL.
`dialect`	`generic`, or `mysql` on a repaired row.
`spider_hardness`	`easy`, `medium`, `hard`, `extra`, or empty.
`extended_hardness`	`easy`, `medium`, `hard`, `extra`, or `extreme`. Always set.
`classification_status`	`classified` or `extended`.
`syntax_status`, `ddl_syntax_status`, `schema_status`	Result of the structural checks.
`syntax_errors`, `ddl_syntax_errors`, `schema_errors`, `classification_errors`	Diagnostic lists. Empty when the check passed.
`status_history`	Ordered record of the pipeline events for the row.
`correction_status`	`corrected` on repaired rows; empty otherwise.
`component1`, `component2`, `others`	Spider counters. Empty when `spider_hardness` is empty.
`validation_source`	Split name and row index in the source copy.

Limitations

Validation and hardness do not execute SQL. A row can be structurally valid and still be a poor answer to its question. The 541 repairs were accepted because the repaired SQL and DDL parse and the names resolve, not because a database returned the expected rows. Columns added during expansion were proposed by openai/gpt-oss-120b. They are not columns from a production schema. Hardness measures statement shape, not how difficult a person would find the question.

License and attribution

The source dataset is released under CC BY 4.0. This refinement is distributed under the same license. Cite the source dataset as well as this refinement.

Citation

@misc{bayer2026sqlcreatecontextrefined,
  title        = {sql-create-context refined},
  author       = {Bayer, Thomas and Grok},
  year         = {2026},
  url          = {https://github.com/bayerth/sql_tools},
  note         = {Refinement of philschmid/sql-create-context-copy, itself a duplicate of b-mc2/sql-create-context.}
}

@misc{bmc2_2023_sqlcreatecontext,
  title        = {sql-create-context Dataset},
  author       = {b-mc2},
  year         = {2023},
  url          = {https://huggingface.co/datasets/b-mc2/sql-create-context},
  note         = {Built by modifying WikiSQL and Spider.}
}

@inproceedings{yu2018spider,
  title        = {Spider: A Large-Scale Human-Labeled Dataset for Complex and Cross-Domain Semantic Parsing and Text-to-{SQL} Task},
  author       = {Yu, Tao and Zhang, Rui and Yang, Kai and Yasunaga, Michihiro and Wang, Dongxu and Li, Zifan and Ma, James and Li, Irene and Yao, Qingning and Roman, Shanelle and Zhang, Zilin and Radev, Dragomir},
  booktitle    = {Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing},
  pages        = {3911--3921},
  year         = {2018},
  address      = {Brussels, Belgium}
}