VISAI-AI/icj-citation-graph

ICJ Citation Graph

This dataset holds the decisions, written pleadings and oral records of the International Court of Justice (ICJ), split into passages with embedding vectors for search by meaning. The documents are linked to the earlier cases, treaties, treaty articles and United Nations resolutions they cite. Where the cited material could be obtained, its text is included too. That covers decisions of the Permanent Court of International Justice (PCIJ) and of other courts and tribunals, treaty articles, and General Assembly and Security Council resolutions together with the citations between those resolutions.

Release release-20261005T012547Z was exported from a single Neo4j snapshot taken on 2026-10-05 and restored into a separate database. Every count in the table and text below comes from a query against that restored copy. The queries and their results are in docs/snapshot-audit.json, and the methods note describes how the export was checked. The counts describe what is stored, including cited cases whose text is not held.

Records	Count
Cases, all courts	571
Documents	6,901
Document passages	1,206,400
Instruments (treaties and similar texts)	370
Provisions (articles of those instruments)	12,375
UN resolutions	20,292
UN resolution passages	169,622
Document-to-case citations	7,450
Document-to-instrument citations	14,763
Document-to-provision citations	10,905
Document-to-resolution citations	1,748
Resolution-to-resolution citations	53,111

Of the cases, 200 are ICJ cases. The rest belong to the PCIJ and other courts. The layer field on documents gives 2,407 ICJ decisions, 2,114 pleadings and 1,875 oral records. The remaining 505 documents come from the PCIJ and other courts and have no layer, and their court field names the court. Filter on layer to keep only ICJ material.

Citations from documents start at ICJ documents only. PCIJ and other-court documents are present as the texts those citations point to, and the release fails its checks if any of them carries a citation of its own. The citation extraction that runs each week reads ICJ decisions. The pleadings and oral records carry treaty citations from an earlier pass that matched treaty names (edge method treaty-name), which is not rerun each week. All 17,464 General Assembly and 2,828 Security Council resolutions have non-empty text.

Files

Format	Location	Contents
Neo4j 5.26 dump	dumps/neo4j.dump	A file that restores the whole graph database, without embedding vectors or vector indexes
Parquet	parquet/	Tables for the dataset viewer, with vectors in the passage, provision, case and resolution tables
Walkthrough	notebooks/dataset_demo.ipynb	Loads the Parquet tables, follows UN citations and searches the vectors
Verification	SHA256SUMS, docs/*.json	File hashes, the snapshot queries and the validation results

In Parquet, a document's citations are the list columns cited_cases, cited_provisions and cited_resolutions on documents, and the resolution citations also have their own edge tables. Citations of a whole instrument rather than one of its articles, citations between provisions, and the judges who sat on each decision are only in the dump.

from datasets import load_dataset

repo = "VISAI-AI/icj-citation-graph"
documents = load_dataset(repo, "documents", split="train")
resolutions = load_dataset(repo, "resolutions", split="train")
edges = load_dataset(repo, "resolution_edges", split="train")

Vectors and row order

The vectors were produced by Qwen3-Embedding-8B with 4,096 dimensions and are stored as float16. They are copied from the snapshot as they are, with nothing recomputed for the export, so new queries must be embedded with the same model and preprocessing to be comparable. docs/validation.json records the observed norm ranges.

The embedding column of chunks, reschunks and provisions holds each row's vector beside its text. cases carries name_embedding and resolutions carries title_embedding, with 427 and 17,650 vectors respectively; a null means the snapshot held no vector for that row. Chunks are sorted by (document_id, seq, id) and resolution chunks by (resolution_symbol, seq, id).

import pyarrow.parquet as pq
batch = next(pq.ParquetFile("parquet/reschunks-000.parquet").iter_batches(
    batch_size=2048, columns=["id", "text", "embedding"]
))
print(batch.schema)

This release uses schema version 3. Chunk tables expose id, the parent ID, seq (position in the document), para (the source's paragraph number, or the speaker in oral records, and empty where neither could be detected), zone (the kind of section the passage came from, such as body, or turn for a speaking turn), layer, text and embedding, and the dump keeps every other graph property. case_id is a string, so zero-padded ICJ IDs and prefixed other-court IDs survive intact. document_count counts the documents linked to a case in this snapshot, and icj_documents repeats that count for ICJ cases only. Some years come from the case metadata snapshot named in manifest.json, and a null year means the year is unknown.

Keeping it current

The graph records its own update progress, so a restored copy can run the weekly update from the code repository and fetch only what is new. An ICJ decision gains un_scanned, ilc_scanned, treaty_scanned and tribunal_scanned once each citation pass has read it. A citation whose target has no text yet is kept on the citing decision in pending_un, pending_ilc, pending_treaty or pending_tribunal, each a list of JSON strings. It becomes an edge once the target's text is fetched. In this snapshot, 2,407 of the 2,407 decisions have been read by every pass, and 5,552 citations are waiting: 5,450 treaty, 102 tribunal, 0 UN and 0 ILC (International Law Commission). A pending citation records what extraction found in the decision and has not been checked against the target.

Restore

Run the graph gives a tested Docker Compose setup that loads the dump, imports the Parquet vectors and points the embedding API at a provider of your choice.

For a manual restore, load the dump into an empty volume while no database is running on it:

docker volume create icj_release_data
docker run --rm -v icj_release_data:/data -v "$PWD/dumps":/backups:ro \
  neo4j:5.26 neo4j-admin database load neo4j --from-path=/backups
docker run -d --name icj-release -p 7474:7474 -p 7687:7687 \
  -e NEO4J_AUTH=neo4j/choose-a-password \
  -e NEO4J_server_memory_heap_max__size=2G \
  -e NEO4J_server_memory_pagecache_size=1G \
  -v icj_release_data:/data neo4j:5.26

The restored database and its indexes need more disk space than the dump file.

The dump was made by copying the snapshot into a fresh database without the three vector properties (embedding, name_embedding, title_embedding). Every label, other property, relationship and application ID was read back and compared. Neo4j's internal IDs are regenerated and differ from the snapshot. The dump was then loaded into an empty database and its node counts matched the snapshot (docs/dump-validation.json, docs/restore-validation.json).

Sources and limits

The ICJ documents combine the CD-ICJ corpus by Sean Fobbe with material retrieved from the ICJ website. Records keep their source metadata where it exists, and some provenance values are file paths from the build machine. The lang label follows the source edition and does not always match every passage.

Text recognition, paragraph boundaries, citation extraction and the matching of cited names to records can all contain errors. A document with no citation edge to a target may still cite it. The validation checks row counts, keys and citation endpoints in full and compares 25 sampled rows per vector table with the stored graph. It does not assess extraction accuracy or legal interpretation.

Terms differ by source; see LICENSE. The dataset authors grant rights only over their own contributions and grant no new rights over the underlying source documents.