Mo7art/Stack2Graph_KG_kotlin

Stack Overflow knowledge graph: Kotlin

This dataset contains Stack Overflow discussions selected by the kotlin programming-language tags using the release's selection registry. The selection refers to programming languages; posts can contain text in different natural languages.

The RDF export is produced by stack2graph.

This Hugging Face dataset provides Parquet tables of RDF quads and a reproducible RDF archive for graph retrieval, SPARQL analysis and retrieval-augmented question answering. It contains the same selected RDF shards, schema and shared documentation as the corresponding Mozilla Data Collective (MDC) publication for the same export identity. The root dataset card and schema image provide the Hugging Face presentation; the archive retains its portable README and exact file inventory.

The programming-language selection is described below. It does not establish the natural languages of post text. The quads configuration exposes the entire graph as the train storage split; it is not an evaluation split or a curated training corpus. Quad counts are not counts of unique questions or question-answer pairs.

Files on Hugging Face

Path	Contents
data/quads-*.parquet	Zstandard-compressed quad tables for the Dataset Viewer and streaming
data/dataset.tar.gz	Reproducible archive containing graphs/ RDF shards, README.md and publication.json
publication.json	The same shard inventory as inside the archive, including stored-file SHA-256 values and export identity
release.json	Archive path, compressed byte count and SHA-256, plus card, inventory and diagram hashes
assets/kg-schema.svg	Schema diagram embedded in this card
README.md	This dataset card with Hugging Face metadata

The RDF paths in publication.json are paths inside the archive. They are not separate downloadable shard paths in this repository. The root release.json identifies the compressed archive; individual RDF hashes identify its extracted files. The export identity is a build fingerprint, not an archive checksum.

Parquet schema and streaming

Each row is one RDF quad. The UTF-8 string columns subject, predicate, object and graph use RDF term syntax: IRIs retain angle brackets, literals retain quotes and language/datatype annotations, and blank nodes retain their identifiers. An empty graph denotes the default graph. Reconstruct N-Quads by joining these terms with spaces and appending .; omit an empty graph term. The original RDF archive preserves the original source bytes.

The generated YAML configuration lists the exact Parquet files for this release, so historical files do not enter the current dataset. Use the repository's full commit SHA to reproduce a release:

from datasets import load_dataset

quads = load_dataset(
    "NAMESPACE/DATASET_REPOSITORY",
    name="quads",
    split="train",
    revision="FULL_COMMIT_SHA",
    streaming=True,
)

Parquet rows preserve graph statements, including duplicates and shared-shard overlap. Their count agrees with the quad count in the source publication inventory. The four columns are RDF terms, not flattened natural-language question/answer fields.

Download and unpack

Download the archive and manifests from the Files and versions tab. For a reproducible download, select a specific repository commit and obtain all files from that same commit. The following Python example uses the official huggingface_hub download API. Replace the repository placeholder with this dataset's ID and the revision with its full commit SHA:

from huggingface_hub import snapshot_download

snapshot_path = snapshot_download(
    repo_id="NAMESPACE/DATASET_REPOSITORY",
    repo_type="dataset",
    revision="FULL_COMMIT_SHA",
    allow_patterns=[
        "README.md",
        "assets/kg-schema.svg",
        "data/dataset.tar.gz",
        "publication.json",
        "release.json",
    ],
    local_dir="dataset-download",
)

Check the downloaded archive's byte count and SHA-256 against release.json, then unpack it into a new directory:

mkdir dataset-unpacked
tar -xzf dataset-download/data/dataset.tar.gz -C dataset-unpacked

Check extracted RDF file sizes and SHA-256 values against publication.json before loading. Preserve the directory layout and named graph IRIs. Load these files into an RDF dataset or quad store; this release is an N-Quads archive alongside the Parquet projection configured for datasets.load_dataset(). RDF loading and SPARQL examples follow below.

Release

Field	Value
Source snapshot	stackexchange_20260630_sakura
Source release date	2026-06
Source archive	stackexchange_20260630/stackoverflow.com.7z
RDF format	N-Quads
RDF shards	658
Quads across all included shards	21597788
Named graphs	41
Stored RDF shard bytes	4717435942
Shard compression	none
Mapping version	kg-map-5-streamed-count-acceptance

The byte count is the sum of the RDF files as stored inside the archive. The size of the downloadable archive also depends on compression and the included metadata.

Contents and selection

The selected language graph is http://stackoverflow.com/kotlin. Schema statements are stored in http://stackoverflow.com/schema. The package also includes the full set of shared tag-documentation shards, including documentation assigned to other language graphs. The counts below cover every included shard.

Named graph IRI	Shards	Quads	Stored bytes
http://stackoverflow.com/.net	1	509,802	106,946,489
http://stackoverflow.com/abap	1	28,199	7,135,975
http://stackoverflow.com/assembly	1	108,506	24,486,120
http://stackoverflow.com/bash	1	269,740	58,415,774
http://stackoverflow.com/c	1	397,463	82,602,742
http://stackoverflow.com/c%23	2	747,572	154,579,782
http://stackoverflow.com/c%2B%2B	1	547,224	115,565,769
http://stackoverflow.com/cobol	1	22,806	5,905,661
http://stackoverflow.com/css	1	377,632	79,878,757
http://stackoverflow.com/dart	1	173,422	38,418,337
http://stackoverflow.com/elixir	1	53,770	13,214,613
http://stackoverflow.com/erlang	1	70,831	17,015,911
http://stackoverflow.com/extra-languages	4	1,946,660	411,030,194
http://stackoverflow.com/f%23	1	99,577	23,061,999
http://stackoverflow.com/fortran	1	69,210	16,302,861
http://stackoverflow.com/go	1	232,066	50,190,743
http://stackoverflow.com/haskell	1	130,034	29,549,858
http://stackoverflow.com/html	1	564,991	118,397,694
http://stackoverflow.com/java	2	852,778	175,779,641
http://stackoverflow.com/javascript	2	835,318	176,973,616
http://stackoverflow.com/julia	1	78,196	18,129,118
http://stackoverflow.com/kotlin	611	8,373,573	1,893,949,876
http://stackoverflow.com/lisp	1	51,116	12,303,281
http://stackoverflow.com/lua	1	121,394	27,520,033
http://stackoverflow.com/matlab	1	192,261	42,328,114
http://stackoverflow.com/objective-c	1	295,433	65,168,125
http://stackoverflow.com/perl	1	199,111	43,997,237
http://stackoverflow.com/php	1	599,966	124,416,158
http://stackoverflow.com/powershell	1	246,683	54,868,728
http://stackoverflow.com/prolog	1	52,218	12,630,513
http://stackoverflow.com/python	2	832,615	172,503,416
http://stackoverflow.com/r	1	309,658	65,272,221
http://stackoverflow.com/ruby	1	332,599	71,344,012
http://stackoverflow.com/rust	1	141,564	31,460,455
http://stackoverflow.com/scala	1	225,800	49,890,758
http://stackoverflow.com/schema	1	246	42,209
http://stackoverflow.com/shell	1	231,941	51,162,808
http://stackoverflow.com/sql	1	385,479	82,063,553
http://stackoverflow.com/swift	1	305,458	65,138,858
http://stackoverflow.com/typescript	1	303,648	66,849,340
http://stackoverflow.com/vb.net	1	281,228	60,944,593

Questions with tags for several programming languages can appear in several language packages. Totals from different packages therefore overlap.

Graph model

Each N-Quads statement contains a subject, predicate, object and named graph IRI. The named graph records the language selection that contains the statement. Entities keep the same IRIs across language graphs, so the same question can be queried in several contexts without introducing a separate question identity. Files are storage shards; splitting a graph across files does not create new logical graphs.

The vocabulary uses the namespace http://stackoverflow.com/schema#, abbreviated as so: below. Its RDF/OWL definitions are included in http://stackoverflow.com/schema.

Questions connect to answers, comments, tags and other questions. The relationship table below gives the direction of each predicate and its exported inverse. An accepted answer is also an answer, and acceptance edges are present only when resolved from the source. Comments can also belong to tag documentation posts. The diagram below shows these entities and relationships.

The diagram's combined "Other question" node illustrates the possible directions of so:duplicateOf. It does not assert that a pair of questions is marked as a duplicate in both directions. Each arrow requires its own source PostLinks record. Neither so:duplicateOf nor so:linkedTo has an automatically exported inverse; an incoming edge is a separate source relation, not a generated reverse of the outgoing edge. An incoming edge is stored in its source question's language graph, so displaying it can require loading another language package.

Timestamps

Timestamps are datatype properties whose objects are xsd:dateTime literals. Each field is a separate RDF statement. In the diagram, rectangular panels group these properties beside their entities; the panels are not RDF entities. Missing source values produce no statement.

Property	Meaning	Applicable entities
so:creationDate	Creation time	Questions, answers, tag documentation and comments
so:lastEditDate	Last edit time	Questions, answers and tag documentation
so:lastActivityDate	Last activity time	Questions, answers and tag documentation
so:communityOwnedDate	Conversion to community wiki	Questions, answers and tag documentation
so:closedDate	Closure time	Questions

The embedded SVG shows representative panels for the question, an answer and a comment. The same rules apply to accepted answers, answer comments and tag documentation as listed above. For example, a question's creation time is represented as the following triple, within its named graph (the value is illustrative):

@prefix so: <http://stackoverflow.com/schema#> .
@prefix xsd: <http://www.w3.org/2001/XMLSchema#> .

<http://stackoverflow.com/q/123>
    so:creationDate "2024-03-15T10:30:00Z"^^xsd:dateTime .

These fields describe the source snapshot, not a complete revision history. Last activity or last edit dates alone do not establish technical freshness.

Entity	Representation
Question	so:Question, identified by http://stackoverflow.com/q/{post_id}
Answer	so:Answer, identified by http://stackoverflow.com/a/{post_id}
Comment	so:Comment, with so:QuestionComment or so:AnswerComment for comments on discussions
Tag	so:Tag, identified by http://stackoverflow.com/tags/{encoded_tag}
Tag documentation	so:TagWiki and so:TagExcerpt, connected to their tag

Questions, answers and documentation posts carry source post IDs. Available properties include titles, original HTML bodies, scores, aggregated upvote and downvote counts, timestamps and content licenses. Comments carry text, scores, creation dates and licenses. Question-specific properties include source answer, view and favorite counts. Missing source values are omitted. Individual voters and user-profile entities are not represented by this mapping.

Relationship	Meaning	Exported inverse
so:hasAnswer	Question to one of its answers	so:answersQuestion
so:acceptedAnswer	Question to its resolved accepted answer	so:acceptedAnswerOf
so:hasComment	Post to a comment on it	so:commentOn
so:taggedWith	Question to one of its source tags	so:tagsQuestion
so:hasTagWiki	Tag to its documentation post	so:tagWikiOf
so:hasTagExcerpt	Tag to its short description	so:tagExcerptOf
so:linkedTo	Directed source PostLinks relation of type 1	None
so:duplicateOf	Directed source PostLinks relation of type 3	None

The listed inverse relationships are written as triples, so querying them does not require an inference engine. A link or duplicate edge does not imply a reverse edge. The mapper also writes applicable superclass types. Answer status uses classes such as so:AnsweredQuestion, so:UnansweredQuestion and so:QuestionWithAcceptedAnswer; missing edges alone do not establish unanswered status. An accepted answer or a high score does not establish technical correctness. The ontology defines so:possibleBetterAnswer, but this exporter does not generate that assessment.

How the language graphs connect

Language membership is determined from question tags using the release's selection registry. Matching covers configured base tags, version patterns and aliases. Answers and discussion comments inherit their question's graph membership; their bodies are not classified independently. Questions without a registry match are routed to the http://stackoverflow.com/extra-languages graph in the full export. This category is not an additional match for questions already assigned to a registered language.

The additional extra-languages dataset

extra-languages can be published as its own dataset alongside the 39 language datasets. Its graph IRI is http://stackoverflow.com/extra-languages. It contains questions with no match in the release's configured language registry, including their answers, comments and outgoing question links. This can include questions about unconfigured languages and questions without a recognized language tag; it is not a classification into one additional programming language.

The graph also holds tags without recorded graph usage and tag documentation without a resolved owner. Its publication package includes its discussion shards, the schema and all shared documentation shards. Regular language packages include shared documentation from this graph, but not its complete discussion collection. Load the extra-languages dataset from the same export identity to resolve links whose target discussions are in that graph. It complements the configured language selections without assigning those discussions to another language.

There are three ways language graphs share or connect content:

1. The same discussion in several graphs. If a question matches both Python and Java, its question, answer and comment statements are emitted into both language graphs with the same entity IRIs. This overlap follows source tags; it is not a translation or an inferred similarity relationship.
2. Shared tags and documentation. Tags retain their identities across graphs. Their definitions, wikis, excerpts and documentation comments are routed to the graphs of questions using those tags. Tags without recorded usage, and documentation without a resolved tag owner, go to extra-languages.
3. Links between different questions. so:linkedTo and so:duplicateOf point from a question to another question's IRI. The edge is stored in the source question's language graphs. Its target can belong to another language graph.

For example, Q1 has tags matching both Python and Java, so its discussion is represented in both graphs. Q2 belongs to Java and is the target of a source link from Q1. The following statements illustrate this arrangement:

Named graph	Subject	Predicate	Object
http://stackoverflow.com/python	Q1	so:linkedTo	Q2
http://stackoverflow.com/java	Q1	so:linkedTo	Q2
http://stackoverflow.com/java	Q2	rdf:type	so:Question

Q1 stands for http://stackoverflow.com/q/1 in both graphs; Q2 stands for http://stackoverflow.com/q/2. The two Q1 entries describe the same entity. There is no additional "same entity" edge. The two so:linkedTo statements are copies of the same source link stored in different named graphs.

These IDs are illustrative. Q2's reference in the Python graph does not by itself include Q2's title, body or answers there. Those descriptions are available when the package containing Q2's discussion has been loaded.

Each language package contains the selected language's discussion shards plus all shared tag-documentation shards and the schema. A Java graph appearing in a Python package's inventory can therefore contain shared documentation without containing Java's full question-and-answer collection. The same applies to extra-languages: its shared documentation can be present without its discussion shards. A referenced question can have no local description when its discussion is only in another package.

To follow these connections across languages, load the relevant packages from the same export identity into one RDF dataset and preserve the named graphs. Shared shards occur in multiple packages; avoid importing them repeatedly into stores that retain duplicate statements. For entity counts across graphs, use COUNT(DISTINCT ?question). Summing quad counts measures graph statements, including repeated statements in different language graphs, rather than unique discussions.

publication.json records each file's path, stored byte count, SHA-256 checksum and graph key, together with the export identity. RDF files are under graphs/. Internal run manifests, QLever indexes, credentials and uploader source are excluded from the archive.

Loading and querying

Load the .nq files as N-Quads and preserve their embedded named graph IRIs. For .nq.gz files, decompress them first if your RDF store cannot read gzip.

The following query counts loaded quads by graph. Use the graph inventory above to check the result, or restrict queries to http://stackoverflow.com/kotlin when working with the selected language.

SELECT ?graph (COUNT(*) AS ?quads)
WHERE { GRAPH ?graph { ?s ?p ?o } }
GROUP BY ?graph

Retrieve questions and their available answers in this language graph:

PREFIX so: <http://stackoverflow.com/schema#>

SELECT ?question ?title ?answer
WHERE {
  GRAPH <http://stackoverflow.com/kotlin> {
    ?question a so:Question .
    OPTIONAL { ?question so:hasTitle ?title }
    OPTIONAL { ?question so:hasAnswer ?answer }
  }
}
LIMIT 25

After loading other language packages, find linked or duplicate questions whose descriptions are available in another named graph:

PREFIX so: <http://stackoverflow.com/schema#>

SELECT DISTINCT ?question ?relation ?target ?targetGraph ?targetTitle
WHERE {
  GRAPH <http://stackoverflow.com/kotlin> {
    VALUES ?relation { so:linkedTo so:duplicateOf }
    ?question a so:Question ; ?relation ?target .
  }
  GRAPH ?targetGraph {
    ?target a so:Question ; so:hasTitle ?targetTitle .
  }
  FILTER (?targetGraph != <http://stackoverflow.com/kotlin>)
}
LIMIT 25

The second query returns only targets with loaded descriptions and titles. A target can occur in several graphs through overlapping tag selections; this does not mean the source and target questions use different languages exclusively.

Uses and limitations

The graph supports SPARQL analysis and retrieval over Stack Overflow discussions, including retrieval-augmented question answering. It represents the source snapshot; later edits and deletions on Stack Overflow are not reflected here.

Quality notes

• Votes are represented as aggregated upvote and downvote counts. Individual vote events and voter identities are not included.
• The graph includes selected moderation-related fields, such as a question's closure date, but does not represent the full moderation context or revision history.
• Community-generated Stack Overflow content can contain errors, outdated information and bias. Changes in languages, libraries and development practices can make advice less applicable over time.
• Scores, accepted-answer status and recent activity are source signals; they do not establish that an answer is correct or current.

License and attribution

This dataset is derived from Stack Overflow content. Its declared dataset license is CC BY-SA 4.0 (CC-BY-SA-4.0). Preserve required attribution and license notices when redistributing derived artifacts, and identify changes where required.

Individual source contributions retain their applicable license terms. Stack Overflow's licensing guidance distinguishes CC BY-SA 2.5, 3.0 and 4.0 according to contribution date. A dataset-level license declaration does not replace those source notices. The graph preserves source-provided license labels in so:contentLicense where available. Consult the original posts and revision records for author attribution and applicable terms. The Hugging Face metadata declares cc-by-sa-4.0, consistent with the MDC publication. For dataset questions, use this repository's Community discussion page; source-post attribution and licensing remain attached to the original content.

Citation

If you use this dataset, cite the Stack2Graph project. The research citation accompanies the required attribution for the underlying Stack Overflow content.

Kleybolte, L. A., Ventura, V., & Zarcone, A. (2026). Stack2Graph: A Structured Knowledge Representation of Stack Overflow Data for Retrieval-based Question Answering. In Proceedings of the Knowledge Graphs and Large Language Models Workshop (KG-LLM) @ LREC26 (pp. 120-133). European Language Resources Association (ELRA). https://doi.org/10.63317/583q4d3mwbdu.

Export provenance

These fingerprints identify the inputs and exporter used for this release. The export identity describes the build configuration and inputs; file checksums are listed separately in publication.json.

Component	Fingerprint
Export identity	4d415cea2977835bf886c177f2bfc309a22738e145591f9c7ae4256e9bd086d5
Source snapshot identity	56fe643952cec74b17ea13d91edc8ad95181326156177c94cce48d3fed9834b9
Language registry	d2a7357c04e7f47d9bb598189aeeb7db0680d6b35f0d9695d8d2e544121afb8e
Ontology	1fc4dc1cc73a1baeddc0e1d70dc22ab95962581c6b61eb895ddcc70956621153
Exporter	8d9b9bac86b5e9c19bb0ce5af022e8a1ecd14c9f8e82e73169cbdb8455470422

Source code and documentation: stack2graph.