anhaidgroup/polaris-wtr-v1

WTR v1

WTR is one of six datasets in [Polaris: Learning to Generate Table Descriptions from Retrieval Feedback](https://arxiv.org/abs/2608.17171), alongside aw, arctic, lter, ecir, and wikitables.

It holds 4,634 tables crawled from web pages and 60 keyword queries over them. For each query–table pair, a person scored how well that table answers that query; those scores are the relevance judgments, and they live in qrels.csv.

The tables have no names. What describes a table is its column names and the text around it on the page, which here is the richest context of the six datasets.

Each Polaris dataset comes in two versions. v1, this one, has the table metadata, the queries, and the relevance judgments. v2 adds the tuples — the contents of each table — and is at polaris-wtr-v2.

Files

`queries.csv` — one row per query. Note that query_id is a plain integer here, not the q1 form used by aw, arctic, lter, and ecir.

query_id,query
1,world interest rates table
2,2008 beijing olympics

`qrels.csv` — one row per query–table pair that a person scored, graded 0, 1, or 2.

query_id,table_id,relevance_score
1,5/1438042988718.8_20150728002308-00068-ip-10-236-191-2_58265756_0.json,2

Judgments came with the source benchmark rather than being made here. They were filtered to the 4,634 tables released, which dropped 4,192 rows — every one of them grade 0, so all 1,271 relevant and 975 highly relevant judgments are intact.

`metadata.csv` — one row per table, here the table above with its column list cut short.

table_id,column_names,table_context
5/1438042988718.8_...json,"[""Country name"", ""2010"", ""2011"", ""2012"", ""2013"", ""2014"", """", """"]","{""pageTitle"": ""Deposit interest rate (%) | Data | Table"", ...}"

table_context holds pageTitle, title, textBefore, and textAfter. table_id contains a slash because it is a crawl path, so inside tuples.zip the file sits at a nested path, Tuples/5/1438...json.csv.

Statistics

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

WTR is the largest of the six by table count, and metadata.csv is 8.8 MB because the page text in table_context is long.

Download

The three files come to about 9.0 MB. You do not need a Hugging Face account to download them.

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-wtr-v1 --repo-type dataset --local-dir wtr

Polaris has six datasets in total and WTR is one of them. Each sits in its own repository, so to download all six quickly — the v1 repositories, without tuples:

for d in aw arctic lter ecir wikitables wtr; do
  hf download anhaidgroup/polaris-$d-v1 --repo-type dataset --local-dir polaris_v1/$d
done

Usage

column_names and table_context are an array and an object, so they need parsing when you load the file. For example:

import ast
import pandas as pd

metadata = pd.read_csv("wtr/metadata.csv")
queries = pd.read_csv("wtr/queries.csv", dtype={"query_id": str})
qrels = pd.read_csv("wtr/qrels.csv", dtype={"query_id": str})

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

relevant = qrels.loc[(qrels.query_id == "1") & (qrels.relevance_score > 0), "table_id"].tolist()

The dtype={"query_id": str} keeps the ids as strings, so the same code works across all six datasets. Keep table_id intact too — the slash is part of the id, not a directory you can drop.

How is this dataset created?

The tables come from a collection of web tables that Chen et al. (SIGIR 2021) built from Common Crawl, a public archive of pages crawled from the open web. The collection pairs 60 queries with roughly 3 million tables and scores each pair 0, 1, or 2: irrelevant, partially relevant, fully relevant. A table scraped from a web page has no title of its own, so what identifies it here is its column names and the text printed around it on the page.

Producing a description for 3 million tables was not practical, so the Polaris authors cut the corpus down. For each query they ran BM25 over the column names and page text, took the 50 highest-scoring tables, and then added back every table already scored above 0.

That leaves the corpus harder than a random sample would, not easier. What survives is every relevant table plus the non-relevant tables that BM25 ranks highest — the ones a system is most likely to confuse them with. That gives 4,634 tables across the 60 queries.

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 benchmark it is built on:

@inproceedings{chen2021wtr,
  author    = {Zhiyu Chen and Shuo Zhang and Brian D. Davison},
  title     = {{WTR}: A Test Collection for Web Table Retrieval},
  booktitle = {Proceedings of the {ACM SIGIR} Conference on Research and Development in Information Retrieval},
  pages     = {2514--2520},
  year      = {2021},
}

License

Contact

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