TrafficSQL-Bench

TrafficSQL-Bench 是一个面向交通领域的 Text-to-SQL / NL2SQL 基准数据集，覆盖机场航空、铁路交通、共享单车、道路车辆、船运物流、赛车竞赛等 6 类交通场景。

本仓库提供可直接使用的数据库 schema、自然语言问题、gold SQL、数据库知识说明，以及逐库压缩后的 SQLite 数据库文件。读者可以按场景、按数据库或按全量任务进行评测。

数据规模

• 数据库：27 个
• 场景：6 个
• Task：1696 条
• Spider 交通子集 task：1135 条
• BIRD train 交通子集 task：561 条
• 完整结果 task：1663 条
• 大结果校核 task：33 条
• 修订后 task：222 条
• 其中修订问题文本：17 条
• 其中修订 gold SQL：218 条
• SQLite 未压缩总大小：约 5.0GB
• 逐库 .7z 压缩包：27 个，压缩后总大小约 258.5MB

Task 版本说明

本仓库中的 tasks/ 是 TrafficSQL-Bench 的正式评测任务。数据整理过程中，部分原始问题文本或 gold SQL 经过人工校核与修订，以减少题意歧义、SQL 明显错误和大结果校核口径不一致等问题。

因此，评测时应以本仓库 tasks/ 目录中的 question 与 gold_sql 为准。为便于追溯，修订过的 task 额外保留了：

• is_corrected
• correction_type
• correction_category
• correction_reason
• original_question
• original_gold_sql

修订类型包括 gold SQL 修正、问题文本修正、大结果校核口径修正，以及 question + SQL 同步修正。当前共 222 条 task 带有修订追踪字段，其中 17 条修订问题文本，218 条修订 gold SQL。

场景

场景 ID	中文场景名	数据库数	Task 数
aviation_airport	机场航空	8	512
rail_train_station	铁路交通	3	84
bike_micromobility	共享单车	2	217
road_vehicle_business	道路车辆	5	388
maritime_shipping_logistics	船运物流	4	262
racing_competition	赛车竞赛	5	233

更完整的数据库与场景清单见：

• docs/DATABASES.md
• docs/SCENARIOS.md

问题示例

共享单车

{
  "task_id": "task_000036",
  "database_uid": "spider_data__bike_1",
  "question": "What are the names and ids of all stations that have more than 14 bikes available on average or had bikes installed in December?",
  "gold_sql": "SELECT T1.name ,  T1.id FROM station AS T1 JOIN status AS T2 ON T1.id  =  T2.station_id GROUP BY T2.station_id HAVING avg(T2.bikes_available)  >  14 UNION SELECT name ,  id FROM station WHERE installation_date LIKE \"12/%\""
}

机场航空

{
  "task_id": "task_001551",
  "database_uid": "train__airline",
  "question": "On August 2018, which day had the highest number of cancelled flights due to the most serious reasons in Dallas/Fort Worth International?",
  "gold_sql": "SELECT T2.FL_DATE FROM Airports AS T1 INNER JOIN Airlines AS T2 ON T1.Code = T2.ORIGIN WHERE T2.FL_DATE LIKE '2018/8%' AND T1.Description = 'Dallas/Fort Worth, TX: Dallas/Fort Worth International' AND T2.ORIGIN = 'DFW' AND T2.CANCELLED = 1 AND T2.CANCELLATION_CODE = 'A' GROUP BY T2.FL_DATE ORDER BY COUNT(T2.FL_DATE) DESC LIMIT 1"
}

道路车辆

{
  "task_id": "task_001257",
  "database_uid": "train__car_retails",
  "question": "Of the clients whose businesses are located in the city of Boston, calculate which of them has a higher average amount of payment.",
  "gold_sql": "SELECT T1.customerNumber FROM customers AS T1 INNER JOIN payments AS T2 ON T1.customerNumber = T2.customerNumber WHERE T1.city = 'Boston' GROUP BY T1.customerNumber ORDER BY SUM(T2.amount) / COUNT(T2.paymentDate) DESC LIMIT 1"
}

铁路交通

{
  "task_id": "task_001161",
  "database_uid": "train__trains",
  "question": "Please list the IDs of all the trains that run in the east direction and have less than 4 cars.",
  "gold_sql": "SELECT T1.id FROM trains AS T1 INNER JOIN ( SELECT train_id, MAX(position) AS carsNum FROM cars GROUP BY train_id ) AS T2 ON T1.id = T2.train_id WHERE T1.direction = 'east' AND T2.carsNum < 4"
}

完整 task 文件见：

• tasks/tasks.jsonl
• tasks/tasks.csv
• tasks/by_database/<database_uid>.jsonl
• tasks/by_scenario/<scenario>.jsonl

目录结构

TrafficSQL-Bench/
  README.md
  docs/
    DATABASES.md
    SCENARIOS.md
    TASKS.md
    OPEN_SOURCE_NOTES.md
  metadata/
    benchmark_summary.json
    databases.csv
    databases.jsonl
    scenarios.csv
    scenarios.json
    database_task_map.csv
  schemas/
    all_schemas.json
    all_schemas.md
    by_database/<database_uid>.json
    by_database/<database_uid>.md
  tasks/
    tasks.jsonl
    tasks.csv
    by_database/<database_uid>.jsonl
    by_scenario/<scenario>.jsonl
  knowledge/
    by_database/<database_uid>_knowledge_database.md
  database_files/
    README.md
    database_files_manifest.csv
    archives_manifest.csv
    archives_manifest.json
    archives/<database_uid>.7z

文件说明

• tasks/tasks.jsonl：全部自然语言问题、gold SQL、schema 引用、执行校核摘要。
• schemas/by_database/：每个数据库的 schema，提供 JSON 与 Markdown 两种格式。
• knowledge/by_database/：每个数据库的业务知识说明，可用于 knowledge-enhanced Text-to-SQL。
• metadata/databases.csv：数据库清单，包括场景、来源、表数、字段数、task 数、SQLite 大小。
• metadata/database_task_map.csv：数据库、task、问题、schema 的映射关系。
• database_files/archives/：每个数据库一个 .7z 压缩包。
• database_files/archives_manifest.csv：压缩包大小、压缩比、SHA256 与包含文件清单。

使用方式

读取全部 task

import json
from pathlib import Path

tasks = []
for line in Path("tasks/tasks.jsonl").open():
    tasks.append(json.loads(line))

print(len(tasks))
print(tasks[0]["question"])
print(tasks[0]["gold_sql"])

解压单个数据库

7z x database_files/archives/spider_data__bike_1.7z -odatabase_layer/spider_data__bike_1

每个压缩包包含：

• .sqlite
• database.json
• knowledge.md
• knowledge_database.md

按场景评测

tasks/by_scenario/aviation_airport.jsonl
tasks/by_scenario/bike_micromobility.jsonl
tasks/by_scenario/road_vehicle_business.jsonl

按数据库评测

tasks/by_database/spider_data__bike_1.jsonl
tasks/by_database/train__airline.jsonl
tasks/by_database/train__shipping.jsonl

Task 格式

每条 task 保留问题、gold SQL、schema 引用和执行校核摘要。示例字段：

{
  "task_id": "task_000001",
  "source": "spider_data",
  "split": "train",
  "db_id": "bike_1",
  "database_uid": "spider_data__bike_1",
  "scenario": "bike_micromobility",
  "scenario_zh": "共享单车",
  "question": "Give me the dates when the max temperature was higher than 85.",
  "gold_sql": "SELECT date FROM weather WHERE max_temperature_f  >  85",
  "schema_ref": "schemas/by_database/spider_data__bike_1.json",
  "knowledge_ref": "knowledge/by_database/spider_data__bike_1_knowledge_database.md",
  "execution_status": "ok",
  "result_mode": "full_rows",
  "result_row_count": 179,
  "is_corrected": false
}

更完整的字段说明见 docs/TASKS.md。

数据来源与使用

TrafficSQL-Bench 基于公开 Text-to-SQL 数据集中的交通相关数据库和问题整理而成，主要包括：

• Spider traffic subset
• BIRD train traffic subset

使用、发布或引用本 benchmark 时，请同时遵守原始数据集的许可证和引用要求。更多说明见 docs/OPEN_SOURCE_NOTES.md。

致谢

感谢 taoyds/spider 和 AlibabaResearch/DAMO-ConvAI BIRD 项目提供高质量的公开 Text-to-SQL 数据资源。TrafficSQL-Bench 的交通领域任务整理工作建立在这些公开数据资源之上。