anhaidgroup/polaris-aw-v2

AW v2

AW is one of six datasets in Polaris: Learning to Generate Table Descriptions from Retrieval Feedback, alongside arctic, lter, ecir, wikitables, and wtr.

It holds 96 tables from a version of AdventureWorks whose column names are cryptic — BusEntId, STRGUID, JobTtl — and 15 keyword queries over them. For each query–table pair, a person decided whether that table answers that query; those decisions are the relevance judgments, and they live in qrels.csv.

Each Polaris dataset comes in two versions. v2, this one, adds the tuples — the rows of each table, in tuples.zip — on top of the table metadata, the queries, and the relevance judgments. v1 is the same without the tuples and is at polaris-aw-v1.

Files

queries.csv — one row per query.

query_id,query
q1,Employee phone number
q2,Discounted products

qrels.csv — one row per relevant pair, here the four tables relevant to q1.

query_id,table_id,relevance_score
q1,HumanResources-2,1
q1,dbo-21,1
q1,Person-11,1
q1,Person-9,1

Judging was done by pooling: for each query, several retrievers proposed candidate tables and a person labelled each one, 1 for relevant and 0 for not. Only the 1s are listed here, so a table missing for a query counts as not relevant.

metadata.csv — one row per table, here two of those four with their column lists cut short.

table_id,table_name,column_names
HumanResources-2,Employee,"[""BusEntId"", ""NationId"", ""Login"", ""OrgLevel"", ""JobTtl"", ...]"
Person-9,PhoneNumberType,"[""PhoneNumTypId"", ""PName"", ""PhoneModDate""]"

Employee holds employee records but has no phone column, and PhoneNumberType holds phone data but nothing about employees. The two connect through BusEntId, which the PersonPhone table spells BusEntID — the same key, two spellings, in one schema.

tuples.zip — one CSV per table, at Tuples/<table_id>.csv. Person-9 is PhoneNumberType, one of the four tables relevant to q1. Its whole contents:

PhoneNumTypId,PName,PhoneModDate
1,Cell,2017-12-13 13:19:22.273
2,Home,2017-12-13 13:19:22.273
3,Work,2017-12-13 13:19:22.273

Each file's header matches that table's column_names. Tables were capped at 500,000 rows; 1 of the 96 hit that cap.

Statistics

A table counts as gold if it scores above 0 for at least one query.

Statistic	Value
Domain	Enterprise
Tables	96
Queries	15
Gold tables	41
Relevant tables per query	min 1, max 8, average 4.0
Metadata fields	table name, column names
Relevance	binary
Tuples	1,543,668 rows across the tables
Rows per table	median 296, min 3, max 500,000
tuples.zip	23.5 MB, 121.7 MB unzipped

Download

The repo is about 23.6 MB, almost all of it tuples.zip, which unpacks to 121.7 MB. You do not need a Hugging Face account to download it.

Option 1 — click the Files tab at the top of this page and save each file.

Option 2 — command line (recommended):

pip install huggingface_hub
hf download anhaidgroup/polaris-aw-v2 --repo-type dataset --local-dir aw

Polaris has six datasets in total and AW is one of them. Each sits in its own repository, so to download all six quickly — the v2 repositories, with tuples, about 4.8 GB in total:

for d in aw arctic lter ecir wikitables wtr; do
  hf download anhaidgroup/polaris-$d-v2 --repo-type dataset --local-dir polaris_v2/$d
done

Usage

column_names is an array, so it needs parsing when you load the file. For example:

import ast
import pandas as pd

metadata = pd.read_csv("aw/metadata.csv")
queries = pd.read_csv("aw/queries.csv")
qrels = pd.read_csv("aw/qrels.csv")

metadata["columns"] = metadata["column_names"].apply(ast.literal_eval)

relevant = qrels.loc[qrels.query_id == "q1", "table_id"].tolist()
# ['HumanResources-2', 'dbo-21', 'Person-11', 'Person-9']

To read one table's tuples without unpacking the archive:

import zipfile

with zipfile.ZipFile("aw/tuples.zip") as z:
    with z.open("Tuples/Person-9.csv") as f:
        rows = pd.read_csv(f)

How is this dataset created?

The tables are the AdventureWorks sample database from Microsoft, a well-known enterprise schema of 96 tables that use many abbreviated names such as BusEntId, STRGUID, and JobTtl. The version used here is the one released with the Columbo paper, where the column names were shortened by hand to simulate the cryptic schemas found in real enterprise data. All 96 tables were kept.

The Polaris authors wrote 15 queries for enterprise data discovery tasks, such as "employee phone number" and "discounted products". They then manually labelled the table–query pairs, giving a binary score of 0 or 1.

Citation

@misc{cai2026polaris,
  title         = {Polaris: Learning to Generate Table Descriptions from Retrieval Feedback},
  author        = {Cai, Ting and Phan, Tuan Minh and Doan, AnHai},
  year          = {2026},
  eprint        = {2608.17171},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  doi           = {10.48550/arXiv.2608.17171},
  url           = {https://arxiv.org/abs/2608.17171}
}

Please also cite the source of the tables:

@inproceedings{columbo,
  author    = {Ting Cai and Stephen Sheen and AnHai Doan},
  title     = {Columbo: Expanding Abbreviated Column Names for Tabular Data Using Large Language Models},
  booktitle = {Findings of the Association for Computational Linguistics: {EMNLP} 2025},
  pages     = {24774--24792},
  year      = {2025},
}

License

What	Comes from	License
Table metadata and tuples	Microsoft's AdventureWorks sample database, with the column names abbreviated by the Columbo authors	MIT, so keep Microsoft's notice with any copy you pass on
Queries and relevance judgments	The Polaris authors	ODC-BY 1.0

Contact

Email minhrua@cs.wisc.edu, valid until May 2029. After that, email anhai@cs.wisc.edu.