BEIR CQADupStack TeX (orgrctera/beir_cqadupstack_tex)

Overview

This release packages CQADupStack / TeX from the BEIR (Benchmarking IR) benchmark as a table-oriented dataset for retrieval evaluation and tooling (e.g. Langfuse-exported runs). The TeX slice is one of the Stack Exchange–hosted sub-corpora in CQADupStack: questions and answers about TeX, LaTeX, ConTeXt, typography, fonts, BibTeX/Biber, and related tooling, drawn from the TeX — LaTeX Stack Exchange community.

CQADupStack was introduced as a benchmark for community question answering (cQA) research. Posts come from multiple Stack Exchange forums; each subforum is distributed as its own subset. Duplicate-question links (and other annotations) support both classification and retrieval-style experiments. In BEIR’s retrieval formulation, queries are question posts and relevant documents are other posts marked as duplicates—so models must retrieve semantically equivalent or overlapping questions from a corpus of prior threads.

BEIR (Thakur et al., NeurIPS 2021) standardizes many heterogeneous IR datasets (lexical vs. semantic gaps, short vs. long documents, domain shift) so sparse, dense, and hybrid retrievers can be compared—including in zero-shot settings where training did not target that forum.

The TeX subforum is a strong fit for retrieval stress-tests: queries mix natural language with macro names, package names, compiler messages, and math/notation, which can challenge both lexical and dense encoders.

This Hub dataset contains 2,906 query-level rows on the test split, aligned with the standard BEIR CQADupStack–TeX evaluation split.

Task

• Task type: Retrieval (document retrieval against an external corpus identified by BEIR document IDs).
• Input (input): The user query text (a TeX/LaTeX Stack Exchange–style question title/body fragment as distributed in BEIR).
• Reference (expected_output): A JSON string encoding the list of relevant document IDs with relevance scores (BEIR qrels: typically binary 1 for relevant pairs), e.g. [{"id": "10943", "score": 1}, ...].
  Evaluators rank a candidate pool (the full CQADupStack–TeX corpus in BEIR) and score overlap with these IDs using standard IR metrics (nDCG, MRR, Recall@k, etc.).
• Metadata: Original BEIR identifiers (query_id) and split name are preserved for traceability.

The retrieval system’s job is to return the correct corpus document IDs for each query when scored against the full TeX forum corpus distributed with BEIR—that corpus is not inlined row-wise in this table.

Background

CQADupStack (original dataset)

CQADupStack is a collection of threads from twelve Stack Exchange communities, built from a 2014 Stack Exchange data dump and annotated for duplicate question detection and related cQA tasks. The TeX subset corresponds to tex.stackexchange.com: practical questions about document classes (e.g. memoir), bibliographies, TikZ/PGF, fonts, math mode, errors/warnings, editors, and compilation workflows—often with code-like tokens and notation that differ from general web text.

Hoogeveen et al. (ADCS 2015) describe the resource and evaluation protocols aimed at comparable training/test splits for retrieval and classification.

BEIR reformulation

BEIR re-hosts CQADupStack per subforum in a shared layout: corpus (JSONL: _id, title, text), queries (JSONL: _id, text), and qrels (TSV: query-id, corpus-id, score). That common format enables cross-dataset benchmarks and zero-shot evaluation of neural retrieval models.

This release

Rows were exported from Langfuse (CTERA AI evaluation pipeline) in a flat, parquet-friendly schema: one row per query with gold relevant document IDs in expected_output for downstream scoring and observability.

Data fields

Column	Type	Description
id	string	Stable UUID for this row in this Hub release.
input	string	Query text (question).
expected_output	string	JSON string: list of objects {"id": "<corpus-doc-id>", "score": <int>} — qrels for that query.
metadata.query_id	string	BEIR query identifier.
metadata.split	string	Split name (here: test).

Splits

Split	Rows
test	2,906
Total	2,906

Examples

Illustrative rows from this release (values taken from the published Parquet; long text may be truncated in the UI).

Example 1 — memoir / table of contents

• input: How to remove double "Contents" heading generated by memoir?
• metadata.query_id: 47225
• metadata.split: test
• expected_output:

[{"id": "10943", "score": 1}, {"id": "50031", "score": 1}, {"id": "3879", "score": 1}]

Example 2 — BibTeX / bibliography updates

• input: References not updating from BibTex
• metadata.query_id: 139375
• metadata.split: test
• expected_output:

[{"id": "61307", "score": 1}, {"id": "105799", "score": 1}, {"id": "151623", "score": 1}]

Example 3 — selective formatting / bold keywords

• input: Is it possible to have certain words in the document always in bold?
• metadata.query_id: 16996
• metadata.split: test
• expected_output:

[{"id": "79592", "score": 1}]

References and citations

BEIR benchmark (aggregation & protocol)

Nandan Thakur, Nils Reimers, Andreas Rücklé, Abhishek Srivastava, Iryna Gurevych. BEIR: A Heterogeneous Benchmark for Zero-shot Evaluation of Information Retrieval Models. NeurIPS 2021 Datasets and Benchmarks Track.

• Paper: OpenReview
• Code: beir-cellar/beir
• Related HF collections under the broader BEIR ecosystem include per-subforum CQADupStack mirrors (naming varies; see BEIR docs for the canonical paths).

@inproceedings{thakur2021beir,
  title={{BEIR}: A Heterogeneous Benchmark for Zero-shot Evaluation of Information Retrieval Models},
  author={Thakur, Nandan and Reimers, Nils and R{\"u}ckl{\'e}, Andreas and Srivastava, Abhishek and Gurevych, Iryna},
  booktitle={Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2)},
  year={2021},
  url={https://openreview.net/forum?id=wCu6T5xFjeJ}
}

CQADupStack (original dataset)

Doris Hoogeveen, Karin M. Verspoor, Timothy Baldwin. CQADupStack: A Benchmark Data Set for Community Question-Answering Research. ADCS 2015.

• Anthology / DOI: 10.1145/2838931.2838934
• Historical resource page (dataset lineage): University of Melbourne scholarly work record

Abstract (short): The authors present a benchmark built from Stack Exchange forums for community question answering, with duplicate annotations and splits designed so that evaluation matches realistic settings (e.g. matching a new post to older threads). The resource supports retrieval and classification baselines with shared preprocessing and metrics—motivating the later BEIR retrieval packaging.

@inproceedings{hoogeveen2015cqadupstack,
  title={{CQADupStack}: A Benchmark Data Set for Community Question-Answering Research},
  author={Hoogeveen, Doris and Verspoor, Karin M. and Baldwin, Timothy},
  booktitle={Proceedings of the 20th Australasian Document Computing Symposium},
  year={2015},
  doi={10.1145/2838931.2838934}
}

Stack Exchange content licensing

Underlying posts are subject to Stack Exchange’s CC BY-SA licensing terms for user-contributed content; respect attribution and share-alike requirements when redistributing or building on the text.