VISAI-AI/icj-citation-graph

ICJ Citation Graph

The International Court of Justice (ICJ) collection provides structured text from decisions, written pleadings and oral records with citation links to cases and legal provisions. It also includes documents from other courts and a United Nations resolution graph containing resolution texts and citation links.

Data

The counts below cover the records included in this release and do not represent the full holdings of each institution. The manifest records table counts and embedding coverage.

Records	Count
Cases, including external cases and citation targets	571
Documents	6,901
Document passages	1,206,400
Legal instruments / provisions	370 / 12,375
UN resolutions / resolution passages	20,291 / 169,617
Document-to-resolution citation links	2,054
Resolution-to-resolution citation links	53,110

The documents comprise 6,396 ICJ records and 505 external-court records. Join documents to cases on case_id to filter by court. UN resolutions include 17,463 General Assembly and 2,828 Security Council records. icj_resolution_edges includes citations from external-court documents too.

Files

Allow approximately 12.15 GB for the download plus additional space for the restored database and indexes.

Files	Contents
dumps/neo4j.dump	Complete Neo4j 5.26 snapshot without embeddings or vector indexes (approximately 0.84 GB)
parquet/	Text, metadata, selected citation relationships and stored embeddings
compose.yaml, runtime/	Graph restore, Parquet vector import and configurable embedding API client

The dump contains all graph relationships including links to judges and countries. Parquet can be used independently of Neo4j and joined to graph records through application IDs because Neo4j internal IDs are regenerated during export.

Get started

Explore the tables

Install datasets and load a table:

from datasets import load_dataset

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

Authenticate with hf auth login if repository access requires it. Available table names appear in the dataset viewer above.

• Notebook: inspect tables, follow UN citations and search a passage sample.
• Run the graph: start Neo4j with Docker Compose and choose how to add vectors. Import them from Parquet or generate them through a hosted or local OpenAI-compatible endpoint.

Start Neo4j with Docker

Install Docker with Compose and allow about 10 GB of free disk space and 4 GB of available RAM for the graph without vectors.

1. Download the runtime files. Install the Hugging Face CLI and log in with a token that can read this repository if it is private.

pip install huggingface_hub
hf auth login
hf download VISAI-AI/icj-citation-graph --repo-type dataset \
  --include "compose.yaml" --include ".env.example" --include "runtime/*" --include "RUN_GRAPH.md" \
  --local-dir icj-graph
cd icj-graph
cp .env.example .env

2. Edit .env. Set NEO4J_PASSWORD to your own password of at least eight characters. For private repository access, also set HF_TOKEN to a token with read access because the download container does not inherit your host login. Leave ICJ_DATA_DIR=./data for this setup.

3. Download the dump and start the graph. Run these commands from icj-graph with Docker running:

docker compose --profile tools run --build --rm fetch
docker compose up --build -d --wait neo4j

The first command downloads and verifies the approximately 0.84 GB dump. The second restores it and starts Neo4j. Neither command calls an embedding API.

4. Open the graph. Visit http://localhost:7474 and log in as neo4j with your password, then run:

MATCH (d:Document)-[:IN_CASE]->(c:Case)
RETURN d, c
LIMIT 25;

This displays documents connected to their cases. Follow RUN_GRAPH.md to add vectors from Parquet or an embedding API. Stop the services with docker compose down when finished; the database remains in its Docker volume.

Embeddings

Parquet includes 4,096-dimensional Qwen3-Embedding-8B vectors rounded to float16 precision and stored as float32 lists.

Tables	Vector column	Availability
chunks, reschunks, provisions	embedding	Every row
cases	name_embedding	427 of 571 rows
resolutions	title_embedding	17,649 of 20,291 rows

Missing vectors are stored as nulls and can be generated through a configured embedding endpoint. The runtime supports different endpoints and models and stores new embeddings in a separate property by default. Stored vectors can be imported without an API call, while new text queries require compatible model and preprocessing settings.

Sources and limits

The ICJ collection combines Sean Fobbe's CD-ICJ corpus with material retrieved from the Court's website. Added material retains source metadata where available. Consult LICENSE before reuse because terms vary by source.

Text recognition, paragraph boundaries and extracted citations may contain errors. Language labels may describe the source edition rather than each passage. Missing citation links do not establish the absence of a citation.

Use SHA256SUMS to verify downloaded files.