Datasets:

misatadata
/

retail-star-answer-key

license: cc-by-4.0
pretty_name: Retail star schema with a verified answer key
language:
  - en
size_categories:
  - 1M<n<10M
task_categories:
  - table-question-answering
tags:
  - synthetic
  - retail
  - star-schema
  - olap
  - clickhouse
  - duckdb
  - text-to-sql
  - benchmark
  - evaluation
configs:
  - config_name: questions
    data_files: answer_key/answer_key.csv
    default: true
  - config_name: orders
    data_files: data/parquet/orders.parquet
  - config_name: order_lines
    data_files: data/parquet/order_lines.parquet
  - config_name: returns
    data_files: data/parquet/returns.parquet
  - config_name: dim_date
    data_files: data/parquet/dim_date.parquet
  - config_name: dim_product
    data_files: data/parquet/dim_product.parquet
  - config_name: dim_customer
    data_files: data/parquet/dim_customer.parquet
  - config_name: dim_store
    data_files: data/parquet/dim_store.parquet

Retail star schema with an answer key

A synthetic retail sales star schema for ClickHouse (1.2 million orders, 3 million order lines, January 2023 to June 2026) and 60 questions, each with its SQL and one checked answer. It tests whether an agent, a cube or a model computes the right number. Read DATASET_CARD.md first: the tables, the measure definitions the answers use, and how the data was made and checked.

Download

The whole package as one file: retail_star_v1.zip (128 MB). The same files, laid out as below, with the Hugging Face CLI:

hf download misatadata/retail-star-answer-key --repo-type dataset --local-dir retail-star

What is in the package

Path	What
`data/parquet/*.parquet`	one file per table, zstd, exact decimals
`data/csv/*.csv`	the same tables as CSVWithNames, `\N` for NULL
`data/ROWCOUNTS`, `data/MANIFEST.json`	rows per table; every file's rows, size and SHA-256
`clickhouse/ddl.sql`, `clickhouse/load.sh`	the tables (MergeTree, pinned types) and a loader that checks row counts
`answer_key/questions.jsonl`	one record per question: the question, the measure definitions it uses, SQL for ClickHouse and DuckDB, the expected value, its tolerance, the traps it tests
`answer_key/answer_key.md`, `answer_key.csv`	the same, as a table
`DATASET_CARD.md`	tables, definitions, as-of date, how it was made and verified
`PATTERNS.md`	the planted patterns and traps, measured; for the person grading, not for agents
`calibration/`	the price calibration: aggregates of US retail listings per kind of product, and the script that makes them
`LICENSE`	CC BY 4.0: use, share and adapt it for any purpose, with credit

Load into ClickHouse

With ClickHouse running in Docker as clickhouse:

CLICKHOUSE_CLIENT="docker exec -i clickhouse clickhouse-client" clickhouse/load.sh parquet

With clickhouse-client on the path, clickhouse/load.sh parquet is enough. clickhouse/load.sh csv loads the CSV files instead. The loader creates the tables, loads each one, and checks every table's row count against data/ROWCOUNTS. It skips tables it has already loaded; --reset drops them and starts over. CH_DATABASE picks a database other than default.

Plain INSERT ... FORMAT Parquet works as well; keep input_format_null_as_default=0 so NULL customer ids and segments stay NULL instead of becoming 0 and ''.

Query without loading

DuckDB reads the Parquet files directly, and every question has DuckDB SQL beside the ClickHouse SQL:

CREATE VIEW order_lines AS SELECT * FROM 'data/parquet/order_lines.parquet';

Grading

Compare an answer with expected within tolerance (questions.jsonl): money exact to the cent, averages of money within ±0.01, ratios within ±0.0001, members by exact name. Every answer was computed on ClickHouse 24.3, 26.8 and 26.9 and on DuckDB and checked equal, and a sample was recomputed without SQL. PATTERNS.md lists what a wrong answer usually looks like, so a failure can be put down to the agent, the model or the dataset.

License

CC BY 4.0 (LICENSE). Credit it as "Retail star schema with a verified answer key, generated with Misata Studio (misata.studio), CC BY 4.0".