VISAI-AI/icj-citation-graph

ICJ Citation Graph

This dataset holds the decisions, pleadings and oral records of the International Court of Justice (ICJ). Each document is split into passages with embedding vectors for search by meaning and linked to the cases, treaties, treaty articles and UN resolutions it cites. Cited material is included where its text could be obtained. This covers decisions of the Permanent Court of International Justice (PCIJ) and other courts, treaty articles and General Assembly and Security Council resolutions with the citations between them.

Snapshot release-20261005T012547Z was taken on 2026-10-05. Counts include cited cases whose text is not held and the methods note explains how they were checked.

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 belong to the ICJ and the rest to the PCIJ and other courts. The layer field splits ICJ documents into 2,407 decisions, 2,114 pleadings and 1,875 oral records. The other 505 documents have no layer and their court field names the court. Filter on layer to keep only ICJ material.

Only the ICJ documents carry citations. PCIJ and other-court documents are cited texts that cite nothing themselves. The weekly extraction reads ICJ decisions only. Pleadings and oral records carry treaty citations from an earlier name-matching pass (edge method treaty-name) that is not rerun. All 17,464 General Assembly and 2,828 Security Council resolutions have 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 sit in the cited_cases, cited_provisions and cited_resolutions columns of documents. Resolution citations also have their own edge tables. Citations of a whole instrument, citations between provisions and the judges on each decision are in the dump only.

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 come from `Qwen3-Embedding-8B` at 4,096 dimensions and are stored as `float16`. They are copied from the snapshot without recomputation, so new queries must be embedded with the same model and preprocessing. `docs/validation.json` records the observed norm ranges.

`chunks`, `reschunks` and `provisions` hold each row's vector in an `embedding` column beside its text. `cases` has `name_embedding` (427 vectors) and `resolutions` has `title_embedding` (17,650 vectors). 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)`.

```python
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, para, zone, layer, text and embedding. seq is the passage's position in its document. para is the source paragraph number or the speaker in oral records and is empty when neither was detected. zone names the kind of section, such as body or turn for a speaking turn. The dump keeps every other graph property.

case_id is a string so zero-padded ICJ IDs and prefixed IDs stay intact. document_count counts a case's documents in this snapshot and icj_documents repeats it for ICJ cases. Some years come from the case metadata snapshot named in manifest.json. A null year means the year is unknown.

Keeping it current

The graph tracks its own update progress, so a restored copy can run the weekly update from the code repository and fetch only what is new. Each ICJ decision gets un_scanned, ilc_scanned, treaty_scanned and tribunal_scanned once the matching citation pass has read it. A citation whose target has no text yet waits on the decision in pending_un, pending_ilc, pending_treaty or pending_tribunal as a list of JSON strings. It becomes an edge once the target's text is fetched.

In this snapshot 2,407 of 2,407 decisions have been read by every pass. There are 5,552 citations waiting: 5,450 treaty, 102 tribunal, 0 UN and 0 ILC (International Law Commission). Pending citations are unverified extraction output.

Restore

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

For a manual restore load the dump into an empty volume with no database 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 is the snapshot copied into a fresh database without the vector properties embedding, name_embedding and title_embedding. Every label, property, relationship and application ID was read back and compared. Neo4j internal IDs are regenerated. The dump was then loaded into an empty database and its node counts matched the snapshot (docs/dump-validation.json and docs/restore-validation.json).

Sources and limits

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

OCR, paragraph splitting, citation extraction and name matching can all contain errors. A missing citation edge does not prove the document never cites the target. Validation checks row counts, keys and citation endpoints in full and compares 25 sampled rows per vector table with the 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 none over the underlying source documents.