Beaver is a holistic framework for evaluating performance on complex, private‑enterprise text‑to‑SQL tasks. This repository includes questions and corresponding annotations. We reserve a portion of the full question set as a private, hidden test set. Each sample contains:

id: ID of the question
category: one of real, complex query, domain-specific query, domain-specific complex query.
- real indicates the query originates from actual query logs. All other categories refer to queries synthesized from templates derived from real queries.
- complex query: queries with high structural complexity (e.g., many joins, nesting) but no domain-specific knowledge
- domain-specific query: queries with low structural complexity but requiring domain-specific knowledge
- domain-specific complex query: queries with both high complexity and domain knowledge
detailed_category: one of real, base, cte, nested, cte-nested, nested-cte. A base query is not treated as a complex query, while a cte, nested, cte-nested, nested-cte query is considered a complex query.
- real indicates the query originates from actual query logs. All other categories refer to queries synthesized from templates derived from real queries.
- base indicates queries synthesized from base templates
- cte indicates queries synthesized from Common-Table-Expression (CTE) templates
- nested indicates queries synthesized from nesting templates
- cte-nested indicates queries synthesized from nesting templates, followed by CTE templates
- nested-cte indicates queries synthesized from CTE templates, followed by nesting templates
contains_domain_knowledge: whether the question includes domain knowledge
db: the ID of the referenced database
question: the natural language question user query
sql: the SQL statement whose execution answers the question
tables: the tables used in the SQL statement
join_keys: the join keys used in the SQL statement
column_mapping: mappings from phrases in the question to specific table columns
domain_knowledge: domain‑specific formatting rules or predicate logic
sub_questions: a decomposition of the question into multiple sub-steps
sub_sqls: the SQL statements corresponding to each sub‑step

Getting started

from datasets import load_dataset
import json

domain = 'dw'
data = load_dataset('beaverbench/beaver-query')
json_fields = ['tables', 'join_keys', 'column_mapping', 'domain_knowledge', 'sub_questions', 'sub_sqls']
for sample in data[domain]:
    sample = {k: (json.loads(v) if k in json_fields else v)  for k, v in sample.items()}
    # print(json.dumps(sample, indent=2))

Citation

@article{chen2024beaver,
  title={BEAVER: an enterprise benchmark for text-to-sql},
  author={Chen, Peter Baile and Yang, Devin and Li, Weiyue and Wenz, Fabian and Zhang, Yi and Tatbul, Nesime and Cafarella, Michael and Demiralp, {\c{C}}a{\u{g}}atay and Stonebraker, Michael},
  journal={arXiv preprint arXiv:2409.02038},
  year={2024}
}