aluha501/community-qa-lab

Community QA Lab v0.5.0

Release date: 2026-08-10. The ML-focused default remains the immutable v0.4.0 API snapshot. v0.5.0 adds the separate technical_qa config from an API snapshot completed at 2026-08-10T02:59:33.978680Z.

What this dataset is

Community QA Lab is a compact answer-reranking dataset. Each row contains one technical question and one candidate answer. Rows with the same question_id form the candidate set for one reranking example.

The dataset measures two signals separately:

• whether the question author accepted the answer;
• how strongly the community voted for the answer.

This makes it possible to evaluate ordinary answer ranking and inspect cases where accepted status and community votes disagree.

Quick start

from datasets import load_dataset

dataset = load_dataset("aluha501/community-qa-lab")
test = dataset["test"]

technical = load_dataset("aluha501/community-qa-lab", "technical_qa")
technical_test = technical["test"]

default is the ML-focused base-data configuration. technical_qa is an opt-in five-site technical-domain benchmark with its own train, validation and test splits. Versioned audit configurations are optional legacy evaluation supplements tied only to the immutable v0.2.0 test set; they join to neither v0.4.0 default nor v0.5.0 technical_qa. All base splits contain answer rows, not one row per question; group by question_id when evaluating.

Inclusion, grouping, and split policy

The public release retains a source thread only when it has at least two candidate answers and exactly one accepted answer. Threads without an accepted answer are excluded; this is an answer-reranking benchmark, not a complete corpus of all community threads.

question_id is a canonical, site-namespaced question identifier. It is the group key for the candidate set, and question groups are disjoint across the three splits. The temporal split is based on question creation time:

Split	Question creation dates	AI	Data Science	Cross Validated	Answer rows
Train	through 2020-12-31	610	1,993	308	7,567
Validation	2021-01-01 – 2021-12-31	82	166	161	957
Test	2022-01-01 onward	192	228	299	1,682

The source/domain mix is therefore not identical across splits. Report results by split and, where possible, by source site.

technical_qa uses a later strict temporal boundary and remains separate from the ML data:

Split	Question creation dates	Software Engineering	Code Review	Quantum	SciComp	CS Theory	Answer rows
Train	through 2022-12-31	365	590	283	34	14	3,294
Validation	2023	76	98	70	6	2	641
Test	2024 onward	85	196	50	6	5	989

Schema

Column	Type	Meaning
question_id	string	Group key for all answers to one question.
question	string	The normalized question text.
answer_id	string	Stable answer identifier.
answer	string	The normalized candidate answer text.
vote_score	int32	Stack Exchange community vote score.
accepted	bool	Whether the question author accepted this answer.
disagreement_type	string	Question-level relationship between accepted status and top community vote; repeated for every candidate row in the group.
updated_at	string	Answer LastEditDate, or CreationDate when it was never edited.
author	string	Display name retained for attribution.
source_url	string	URL of the original answer.
license	string	Per-answer content_license from the archived API response.

The model input is normally only question and answer. The remaining columns are labels, grouping information, timestamps, or attribution.

Disagreement labels

• accepted_unique_top: the accepted answer has the unique highest vote score.
• accepted_tied_top: the accepted answer shares the highest vote score.
• accepted_below_top: another answer has a higher vote score than the accepted answer.

accepted and vote_score are separate evaluation signals. Neither is a guarantee of factual correctness or answer quality. disagreement_type is a label/analysis field, not a model input.

Release contents

Split	Answer rows	Questions
Train	7,567	2,911
Validation	957	409
Test	1,682	719
Total	10,206	4,039

The release refreshes 4,367 seed question IDs through the Stack Exchange API. After current answer/acceptance/tag checks and strict attribution filtering, 4,039 groups remain from ai.stackexchange.com, datascience.stackexchange.com, and stats.stackexchange.com. The seed pool combines the immutable v0.2.0 IDs with eligible local crawl IDs; it is not an exhaustive enumeration of every ML question on these sites.

Disagreement totals across the 4,039 questions are:

Label	Questions
accepted_unique_top	2,621
accepted_tied_top	830
accepted_below_top	588

technical_qa contents (v0.5.0)

Split	Answer rows	Questions
Train	3,294	1,286
Validation	641	252
Test	989	342
Total	4,924	1,880

The 3,655 seed IDs came from crawler checkpoints for five selected sites. An API refresh retained 1,930 groups with one accepted answer, at least two answers and a configured tag; 50 more were excluded because question-level license attribution was missing. The seed inventory is not exhaustive.

Disagreement totals are 1,165 accepted_unique_top, 346 accepted_tied_top, and 369 accepted_below_top groups. All 4,924 published answer rows carry CC BY-SA 4.0 in the archived API response.

Benchmark release history

v0.5.0 (technical_qa config; default unchanged)

v0.5.0 adds a separate technical-domain config rather than diluting the ML-focused default benchmark. The snapshot has 1,880 groups, with a 342-group test set beginning in 2024 and all five sites represented in every split.

Test model	Accepted@1	Accepted MRR	Vote@1	Vote nDCG
Random	0.3684	0.6408	0.4532	0.8191
Answer length	0.5526	0.7494	0.5409	0.8569
BM25	0.5205	0.7381	0.5322	0.8557
BGE-small	0.4357	0.6870	0.4825	0.8368
MS MARCO MiniLM cross-encoder	0.3743	0.6462	0.4474	0.8265

BGE-small and the cross-encoder significantly trail BM25 on accepted-answer Hit@1/MRR under paired bootstrap. Answer length is the strongest overall baseline, and 262/342 test groups are length-skewed. These are weak-label and domain-shift findings, not evidence that answer length determines factuality.

v0.4.0 (current base data and benchmark)

v0.4.0 preserves the compact 11-column schema while adding Cross Validated, refreshing every retained post through the Stack Exchange API, and increasing the test split from 127 to 719 question groups. The strict temporal boundaries are train through 2020, validation in 2021, and test from 2022 onward. The test release gate requires at least 500 groups.

The exact API response envelopes are archived as a deterministic compressed sidecar and listed by SHA-256 in the release manifest. Threads are excluded if the current API response lacks two answers, one accepted answer, a configured ML tag, or explicit per-post licensing. In particular, 175 otherwise eligible groups were conservatively excluded because question-level content_license was absent; no license was inferred.

Test model	Accepted@1	Accepted MRR	Vote@1	Vote nDCG
Random	0.4687	0.7191	0.5772	0.8711
Answer length	0.6064	0.7924	0.6676	0.9020
BM25	0.5925	0.7830	0.6718	0.9031
BGE-small	0.5647	0.7679	0.6662	0.8983
MS MARCO MiniLM cross-encoder	0.5661	0.7691	0.6704	0.8997

Both pretrained models trail BM25 on Accepted@1, and their paired 95% bootstrap intervals include no improvement over BM25. Answer length is also a strong target-correlated baseline. This is an important behavioral-bias result: the dataset should be used to study weak-label ranking and disagreement, not to claim that generic semantic models recover factual truth.

v0.2.1 (benchmark/code; base Parquet unchanged)

This release adds reproducible CPU reports for random, answer-length, BM25, BAAI/bge-small-en-v1.5, and cross-encoder/ms-marco-MiniLM-L-6-v2. Each run stores the input hash, model revision, package versions, prediction hash and runtime. Evaluation adds question-level bootstrap confidence intervals, per-site macro reporting, paired model comparisons, and a reporting-only answer-length spread slice.

Test model	Accepted@1	Accepted MRR	Vote@1	Vote nDCG
Random	0.4724	0.7214	0.5354	0.8631
Answer length	0.6142	0.7982	0.7323	0.9275
BM25	0.5276	0.7552	0.6535	0.9101
BGE-small	0.6299	0.8091	0.6772	0.9112
MS MARCO MiniLM cross-encoder	0.5591	0.7690	0.5906	0.8841

These remain behavioral metrics over accepted status and votes. They do not establish factual answer quality. The paired comparisons and 95% confidence intervals are part of the versioned benchmark artifacts.

v0.3.0 (audit supplement; separate configuration)

audit_v0_3_0 is a separate test-only configuration with 290 answer-level rows joinable to the immutable v0.2.0 test split by (question_id, answer_id). At publication it did not change the 11-column v0.2.0 default files; it is not joinable to the current v0.4.0 default.

The audit is a model-assisted manual audit: candidate order was deterministically randomized and the reviewer did not see accepted status, vote score, disagreement label, answer ID, source URL, or the Stack Exchange acceptance/voting interface until group decisions were frozen. Candidate choice prioritized factuality, then relevance, then helpfulness. A small number of technical decisions record official papers or documentation; all other rows are explicitly text_only or not_verifiable.

Audit field	Meaning
group_decision	one_preferred, tied_preferred, no_preferred, or insufficient_evidence; repeated for the candidate rows in its question group.
is_preferred	Whether this candidate was selected by the blinded audit. It is false for all candidates in abstained/no-preference groups.
factuality, relevance, helpfulness, confidence	Categorical comparative assessment fields; they are not expert truth labels.
evidence_status, evidence_urls	Whether a recorded non-Stack-Exchange external source informed the assessment. source_checked always has one or more URLs.
rationale	Non-empty comparative audit rationale for this candidate.
reviewer_type, audit_protocol_version, audited_at	Fixed provenance fields for this release.

The audit has 122 groups with one preferred answer, four no_preferred groups, and one insufficient_evidence group. Its preferred signal agrees with the accepted answer in 71/122 auditable groups and with a top-voted answer in 76/122; this is descriptive agreement, not a claim that either signal is ground truth.

Audit labels are evaluation-only, model-assisted, non-independent, and contaminated by their use of the public question/answer text. Do not train, prompt-tune, select, or otherwise tune a model against these labels on the same public test set. See docs/audit_protocol_v0.3.0.md and audits/v0.3.0/manifest.json for the complete protocol and immutable base test hash.

v0.3.1 (preferred audit supplement)

audit_v0_3_1 preserves every v0.3.0 group_decision and is_preferred value, but corrects the candidate-level assessment expansion. It is explicitly a post-unblind quality revision, not a fresh blind or independent audit.

In v0.3.1, all text-only candidates use factuality=not_verifiable; only 17 rows with recorded external sources retain supported or partially_supported factuality. Every candidate has a distinct rationale grounded in its answer text. The mechanical quality report records zero duplicate rationales, zero legacy generated prefixes, zero short rationales, and no assessment field that deterministically encodes is_preferred.

Use audit_v0_3_1 instead of v0.3.0 for candidate-level analysis. Audit Hit@1/MRR are unchanged because the preference target is deliberately frozen. See docs/audit_protocol_v0.3.1.md, audits/v0.3.1/manifest.json, and audits/v0.3.1/quality_report.json.

Provenance and licensing

v0.4.0 and v0.5.0 use Stack Exchange API v2.3 with the pinned withbody filter. The builder stores each raw response envelope immediately, honors API backoff, resumes from matching request caches, and archives the exact response files into a deterministic source/api_snapshot.jsonl.gz sidecar (177 responses for v0.4.0 and 102 for technical_qa). Seed, snapshot, source-archive and Parquet SHA-256 values are recorded in each release manifest and verified before publication. Every answer row retains its source URL, author and per-record ContentLicense.

The data contains 2,017 CC BY-SA 3.0 rows and 8,189 CC BY-SA 4.0 rows. Do not replace the per-record license values with one blanket license. Users are responsible for complying with the applicable Stack Exchange terms and the license attached to each record.

Limitations

Accepted status is a user interaction signal and vote score is a community interaction signal. They are useful weak labels for ranking analysis, but they are not human factuality judgments. Because the release requires one accepted answer and at least two candidates, it does not represent questions with no accepted answer. Answer length is strongly predictive of these behavioral targets in v0.4.0, so improvements must be compared against the length baseline and reported on length_matched as well as length_skewed groups.

The benchmark is English-only. default covers three ML/statistics-oriented communities; technical_qa covers five selected technical communities but is dominated by Code Review in its test split. Both collections are seed-based, not exhaustive, and site activity/domain mix changes over time. The public origin also creates substantial language-model contamination risk: strong results may partly reflect pretraining exposure. Exact normalized text has zero overlap across the three v0.4.0 splits, but that check does not rule out near-duplicate semantics or external pretraining contamination.

The audit_v0_3_0 and audit_v0_3_1 configurations remain available for historical reproducibility only. They cover the 127-group v0.2.0 test split and must not be joined to or presented as an audit of v0.4.0 default or v0.5.0 technical_qa.