Versioned open data

Download the graph.

Explore a fixed snapshot of the news-derived Causals Graph. Each file is structured for joining nodes, relationships, and source-news records. A graph link alone does not establish causation.

Current snapshot

v0.1.0 2026-10-09

View version on GitHub ↗

369,184 nodes · 1,272,418 directed edges · 41 observed relation codes. This version is a fixed snapshot from 9 October 2026; Explore displays a separately updated graph, so its counts can differ. You can also browse the snapshot on Hugging Face ↗.

Start with a small sample. The 10-minute notebook ↗ uses 62 nodes and 9 edges from this version to join records and inspect source-news IDs. Run it in Colab ↗ or browse the sample ↗. The sample illustrates file structure; it is not a representative accuracy test.

nodes.jsonl.gzNodes · gzip-compressed JSON Lines · 12.2 MB edges.jsonl.gzDirected edges · gzip-compressed JSON Lines · 60.0 MB manifest.jsonCounts, relation distribution, and SHA-256 checksums

Browse on Hugging Face ↗Read the data card ↗Read the license ↗

Data structure

Inside the files.

Both data files are gzip-compressed JSON Lines: one JSON record per line. Match an edge's source_id and target_id to a node's id.

Nodes nodes.jsonl.gz

id: The node's identifier in this snapshot.
kind: Broad record kind: entity, news, or other.
labels: One or more classes, such as Entity and Company, or Episodic for a source-news record.
name: A display name where present. Some nodes have no name in this file.
symbol, valid_at, published_at: Symbol and time metadata. Values can be null.

Edges edges.jsonl.gz

id: The edge's identifier in this snapshot.
source_id, target_id: Directed endpoints; both match node IDs in the same snapshot.
type: Top-level edge type: MENTIONS, RELATES_TO, or VARIANT_OF.
relation: Specific code for a RELATES_TO edge; otherwise the same value as type. This snapshot contains 41 observed codes, including descriptive and causal-claim relationships.
source_episode_ids: IDs of source-news nodes in the node file. The list may be empty.
valid_at, invalid_at, as_of: Time metadata attached to the edge. Values can be null.

See the observed node labels and relation codes in v0.1.0

Node labels

Entity, Episodic, Event, Company, Organization, Location, Concept, Security, Software, CreativeWork, Infrastructure, Commodity, FinancialIndex, EventClass, Currency.

Relation codes

AFFECTS, AFFILIATION, CLASSIFIED_AS, COMPETES_WITH, CONSTITUTES, CONSUMES, CREATED_BY, EXPORTS_TO, EXPOSED_DOWN, EXPOSED_UP, GUARANTEES, HAS_SECURITY, INDEX_CONSTITUENT, INPUT_TO, INSTANCE_OF, INVESTS_IN, INVOLVES, ISSUED_BY, LOCATED_IN, MENTIONS, NOT_TRACKING, OPPOSES, OTHER, OWES_TO, OWNS, PART_OF, PRODUCES, REGULATES, SHIFTED, SHIPS_VIA, SPORTS_MATCH, STATED_CAUSE, STRIKES, SUPPLIES, SUPPORTS, TRACKS, TRANSMITS, TRIGGERED, VARIANT_OF, WORKS_FOR, WORK_RIGHTS_OF.

These are values observed in this version, not a promise that every code has the same meaning in every future version. For counts by relation, see manifest.json ↗.

The manifest.json file also lists record counts and SHA-256 checksums. For interpretation and quality notes, read the data card ↗.

License and citation

Reuse with attribution.

This snapshot is available under CC BY 4.0. Credit Causals Lab, identify the version, link to the source and license, and indicate changes. The license applies to these dataset files; material you combine with them may have its own terms.

Suggested citation: Causals Lab. (2026). Causals Graph (v0.1.0-2026-10-09) [Data set]. Version archive.

Corrections and removals
Data can be wrong or incomplete. Send correction or removal requests privately to data@causals.org. Please do not put personal or sensitive details in a public GitHub issue.