pretty_name: Text2SQL-Decisions Benchmark Seed
license: cc0-1.0
language:
- en
task_categories:
- text-classification
- text-generation
tags:
- text-to-sql
- evaluation
- synthetic
size_categories:
- n<1K
configs:
- config_name: default
data_files:
- split: test
path: test.jsonl
- config_name: ecommerce_e2e
data_files:
- split: test
path: ecommerce-e2e.json
Text2SQL-Decisions Benchmark
English evaluations for choosing SQL plans and answering database questions. This repository contains a 24-question synthetic test, 20 application-level questions, and model results on the separate Text2SQL-Decisions dataset.
Model results
Each model received the same question, database schema, and four SQL candidates. Reference answers were excluded from model input. Accuracy measures selection of the correct candidate.
| Model | Test accuracy (2,000) | Full dataset accuracy (25,000) |
|---|---|---|
| D1 | 100.00% | 99.996% |
| Clef Flash | 98.35% | 98.46% |
| D1 Omni 600M | 47.20% | 46.528% |
The 2,000-question test split is the primary comparison. The full dataset includes 20,000 training, 2,000 development, 1,000 calibration, and 2,000 test questions; its score is a dataset audit. Results measure four-choice selection, not free-form SQL generation.
D1 identifies itself as liquid/d1-20260930; equivalence to D1 3B is unverified. Omni's deployed weights revision and precision are unknown. Its public encoder truncates long options, but whether that affected this deployment remains unverified.
Reports contain exact counts, dataset checksums, routing details, and scores by split and source: D1, Clef Flash, D1 Omni 600M, comparison. All three runs completed 25,000 questions. API errors were excluded from scoring and retried. Reported costs cover successful responses; missing costs remain unknown.
Synthetic SQL test
The 24 questions cover purchases/refunds, inventory, and subscriptions, with eight per schema. They test NULL handling, joins, aggregation, ordering, date boundaries, and window functions. Four executable candidates return distinct results; correct answers occupy each position six times.
Clef Flash scored 24/24. Questions and reference answers were generated by the same assistant and checked against expected database results, without independent human review. This small test supports limited conclusions about broader SQL ability.
Application test
The 20 Olist questions exercise plan selection, PostgreSQL execution, and final answers. Of 18 answerable questions, 12 received correct direct answers, four received incorrect answers, and two returned alternatives containing the correct answer. Both unanswerable questions were refused as expected. Direct-answer accuracy was 66.67%.
One delivery-date reference omits a delivered-status filter that the application adds, leaving an interpretation dispute. Conversation chains and full meter coverage remain outside this test. See results.
Usage and evaluation splits
from datasets import load_dataset
choice = load_dataset("Chulinz/Text2SQL-Decisions-Benchmark", "default", split="test")
ecommerce = load_dataset("Chulinz/Text2SQL-Decisions-Benchmark", "ecommerce_e2e", split="test")
For application evaluation, set up the harness and run:
npm run test:live -- --cases /absolute/path/ecommerce-e2e.json
fixture.sql and build-benchmark.py reproduce the synthetic cases. manifest.json records frozen checksums; overlap-audit.json checks exact overlap with the source dataset. Unknown overlap with model pretraining data limits interpretation.
For future fine-tuning, use the source dataset's 20,000 training rows. Reserve development for model selection, calibration for probability adjustment, and test plus this benchmark for evaluation. No fine-tuning has been performed.
License
Authored questions, SQL, fixtures, and documentation use CC0-1.0. Olist records are excluded and require separate access under CC BY-NC-SA 4.0. Model and external-data terms remain separate.