birdsql/bird_mini_dev

BIRD SQL Mini-Dev

2026 update

This update provides 500 aligned question IDs for SQLite, MySQL and PostgreSQL. It revises the annotations from the last update after community feedback and a systematic, human-in-the-loop review of questions, evidence and gold SQL.

Our aim is a benchmark that requires an agent to understand the request, explore the permitted database and construct the query. We clarified materially ambiguous output contracts, corrected reference SQL and replaced selected instances, while removing evidence that supplied unnecessary schema links, query plans or answers. We carefully considered each suggestion against the question's intended meaning and the underlying database, and incorporated changes where the evidence supported them.

Moving forward

BIRD 1.0 achieved broad adoption and helped shape text-to-SQL research and industry development by establishing standardized task formats and realistic database environments, unlocking technological advances in RAG, multi-agent systems, and knowledge generation. Building on this success and sustained community enthusiasm, we are developing BIRD 2.0 to extend evaluation beyond text-to-SQL into developing, testing, and maintaining real-world data workflows. Our goal is to better measure frontier models' data intelligence and their reliability in realistic environments, including enterprise and personal use.

Follow the BIRD website for updates on our next-generation benchmarks: LiveSQLBench 2.0, BIRD 2.0, and DII v1.0.

All contributions, real-world samples, and suggestions of any kind are welcome; get in touch with the BIRD Team (bird.bench25@gmail.com) to collaborate.

We sincerely thank OpenAI, Amazon, and Google Cloud for their early support and collaboration.

Repository layout

Directory	Contents
data/	Current annotations for all three SQL dialects
baseline_inference/	Ready-to-use baseline prompts and supporting schemas
archive/	Historical 2025 release

Interactive tasks are hosted separately on Harbor Hub.

Load the current data

from datasets import load_dataset

mini_dev = load_dataset("birdsql/bird_mini_dev")
print({name: len(split) for name, split in mini_dev.items()})
# {'mini_dev_sqlite': 500, 'mini_dev_mysql': 500, 'mini_dev_pg': 500}

Each split contains 500 instances, IDs 0–499. Question and evidence text match across dialects; reference SQL is dialect-specific. Match records by question_id and dialect.

Baseline results

Fixed-prompt baseline execution accuracy (EX, %) on 500 questions per dialect. Model rows are sorted by Overall EX, from highest to lowest.

Model	SQLite	MySQL	PostgreSQL	Overall
Empirical Upper Bound	99.60	93.80	97.00	96.80
GPT-6.1 Sol	88.00	84.80	89.00	87.27
Claude Opus 5.5	90.00	84.40	87.00	87.13
Claude Sonnet 5.5	86.60	80.40	85.60	84.20
Qwen3.8-Flash-Next	70.20	67.00	67.80	68.33
DeepSeek-V4.1-Flash	61.00	55.80	62.20	59.67
Claude Haiku 5.5	59.60	55.20	58.60	57.80
GPT-6 Luna	54.00	51.00	51.80	52.27
GLM-5.3	51.40	48.40	48.40	49.40
Qwen3.6-35B-A3B	21.20	22.20	20.00	21.13

Overall is EX across all 1,500 dialect-specific answers, equivalent to the mean of the three dialect scores. Empirical Upper Bound counts a question as correct within each dialect if at least one evaluated model answers it correctly.

Baseline and Harbor

• Baseline: use each row's prompt in baseline_inference/ as model input. The separate gold SQL field is for evaluation only. See the baseline guide.
• Harbor: use BIRD Mini-Dev on Harbor Hub for interactive evaluation. Submit the final query to /app/answer.sql. The Harbor dataset includes the local isolation provider; solver and verifier environments have network access disabled.

Database images are available for SQLite, MySQL, and PostgreSQL, with amd64 and arm64 support.

Acknowledgements

Thank you to the BIRD community for careful reports on question wording, evidence, gold SQL and database consistency. We retain our thanks to the NVIDIA researchers and Thinking Machines Lab for their detailed benchmark reports and discussions, and to GitHub community reporters @ivermin1123, @christian-gheorghe, @headinthebox, @tejusarora, @gkorland, @tohpinren, and @MichaelBruwerSwartz.

Previous release and citation

The original 2025-07-04 release remains in archive/bird_mini_dev_2025-07-04/. Its annotations and baseline tables are historical, not current results.

@article{li2024can,
  title={Can llm already serve as a database interface? a big bench for large-scale database grounded text-to-sqls},
  author={Li, Jinyang and Hui, Binyuan and Qu, Ge and Yang, Jiaxi and Li, Binhua and Li, Bowen and Wang, Bailin and Qin, Bowen and Geng, Ruiying and Huo, Nan and others},
  journal={Advances in Neural Information Processing Systems},
  volume={36},
  year={2024}
}