Madeleinex/chicken-graph

ICJ Citation Graph

The International Court of Justice's public record as a graph: 199 of the Court's 200 cases, from Corfu Channel in 1947 to 2026, in three layers.

• Decisions (2,407): judgments, orders, and advisory opinions, plus the separate and dissenting opinions the judges append.
• Written pleadings (2,114): the memorials, counter-memorials, applications, and declarations of intervention the parties filed, in 194 of the cases.
• Oral records (1,875): verbatim hearing transcripts, in 163.

6,396 documents in all, parsed into structured text and split into about 1.1 million searchable passages. Every passage carries a Qwen3-Embedding-8B vector (open-weight, 4096-dim, unit-normalized) for semantic search. The decision layer also has the full citation graph: each decision linked to its case, the earlier cases it cites, the provisions it invokes, and the judges who sat.

Which form to grab

Form	Files	Layers	Needs
Graph, decisions	dumps/icj-decisions.dump (4.2 GB)	decisions	Neo4j 5.26
Graph, full	dumps/icj-full.dump (13.7 GB)	all three	Neo4j 5.26
Portable, decisions	dataset/ (JSONL + npy, 1.0 GB)	decisions	numpy for vectors
Portable, all layers	dataset_layered/	all three, filter by layer	numpy for vectors
Columnar, all layers	parquet/ (8.3 GB)	all three, vectors inline	anything that reads Parquet

Embeddings ship in every form: node properties in the dumps, one embeddings_<layer>.npy per layer in the portable sets, an embedding column in Parquet. Nothing needs re-embedding. Embed your queries with the same model or the scores are meaningless.

Quick start

from datasets import load_dataset
docs   = load_dataset("Madeleinex/chicken-graph", "documents", split="train")
chunks = load_dataset("Madeleinex/chicken-graph", "chunks", split="train")

Parquet is the format this page's viewer and SQL console read natively, and the one form where the vectors are a real column rather than a sidecar:

import duckdb
duckdb.sql("""
  SELECT case_id, title, para, text
  FROM 'parquet/chunks-*.parquet' c JOIN 'parquet/documents.parquet' d
    ON c.document_id = d.id
  WHERE c.layer = 'pleadings' AND text ILIKE '%provisional measures%'
  LIMIT 10
""")

case_id is a zero-padded string ("001"). Parquet preserves that; a JSONL reader that infers types will turn it into an integer and drop the padding, so pass dtype={"case_id": "string"}.

A runnable walkthrough (load, join, reconstruct a document, semantic search) is in notebooks/dataset_demo.ipynb.

Portable files

dataset/ holds the decision layer; dataset_layered/ holds all three with a layer field (decision / pleadings / oral) on every documents.jsonl and chunks.jsonl row. Filter to any subset in one line:

import pandas as pd
docs = pd.read_json("dataset_layered/documents.jsonl", lines=True, dtype={"case_id": "string"})
pleadings = docs[docs.layer == "pleadings"]   # 2,114 filings; oral -> 1,875

cases.jsonl (199 rows)

{"case_id": "070", "name": "Military and Paramilitary Activities in and against Nicaragua (Nicaragua v. United States of America)", "year_start": 1984, "year_end": 1991, "icj_documents": {"decision": 8, "pleadings": 8, "oral": 3, "other": 4}}

A null year_end means the case is still pending (22 rows, exactly the ICJ's official pending list). icj_documents is a coverage snapshot of how many documents of each kind the ICJ site publishes for the case, scraped 2026-07-14; it describes what exists upstream, not what is in this release.

provisions.jsonl (356 rows)

{"key": "statute:article_36", "title": "Article 36", "level": "article", "instrument": "statute", "text": "1. The jurisdiction of the Court comprises all cases ..."}

text is the full article body, present on 346 of 356 (empty on chapter containers and a few keys pointing outside the four parsed instruments: UN Charter, ICJ Statute, Rules of Court, Practice Directions).

documents.jsonl (2,407 decision rows; 6,396 in dataset_layered/)

{
  "id": "ICJ_011_USNationalsMorocco_FRA_USA_1952-08-27_JUD_01_ME_00_EN",
  "case_id": "011",
  "title": "RIGHTS OF NATIONALS OF THE UNITED STATES OF AMERICA IN MOROCCO (FRANCE v. UNITED STATES OF AMERICA)",
  "document_type": "judgment", "lang": "EN", "issued_date": "1952-08-27",
  "applicant": "FRANCE", "respondent": "UNITED STATES OF AMERICA", "stage": "merits",
  "cited_cases": ["007", "008"],
  "cited_provisions": ["statute:article_38", "statute:article_40"]
}

id, case_id, title, document_type, lang, filename, cited_cases, cited_provisions are on every row. The rest appear when the document states them: issued_date on 1,429, applicant/respondent on 574/572 (raw header strings, often ALLCAPS; normalized country identities live in the graph's Country nodes). cited_cases and cited_provisions are the citation edges, so the whole graph is reconstructable from these files alone.

chunks.jsonl + embeddings_<layer>.npy

{"id": "166-20240131-jud-01-00-en:0", "document_id": "166-20240131-jud-01-00-en", "seq": 0, "para": null, "zone": "title", "text": "..."}

Chunks are cut on the unit each document uses: a pleading breaks at the paragraph number counsel gave it, a hearing at each speaker's turn (that label is in para; oral chunks are marked zone: "turn"). Legacy scans with no detectable structure (about 24% of pleadings chunks, 17% of oral) fall back to a fixed window with para empty.

Each .npy is a float16 matrix, L2-normalized (so cosine similarity is a dot product). Row i is the i-th chunk of that layer in file order; the decision slice is byte-compatible with dataset/embeddings.npy.

import json, numpy as np
emb = np.load("dataset/embeddings.npy", mmap_mode="r")   # (121438, 4096) float16
chunks = [json.loads(l) for l in open("dataset/chunks.jsonl")]
scores = emb.astype(np.float32) @ query_vector           # query embedded with Qwen3-Embedding-8B

Ready-made graph (dumps/)

Two dumps: icj-decisions.dump (decisions) and icj-full.dump (all three layers). Both carry the chunk_text full-text index (all layers) and case_name_embedding vector index. icj-decisions.dump also ships its chunk_embedding vector index ready; icj-full.dump holds every chunk's vector as a property but leaves that index for you to CREATE on restore, since an HNSW over a million 4096-d vectors needs more memory than most machines have (or use the .npy with FAISS).

# load into a fresh volume (neo4j-admin loads <db>.dump, so name it neo4j.dump first)
cp dumps/icj-full.dump neo4j.dump        # or icj-decisions.dump for decisions only
docker run --rm -v icj_data:/data -v "$PWD":/backups:ro neo4j:5.26 \
  neo4j-admin database load neo4j --from-path=/backups --overwrite-destination=true

docker run -d --name icj-neo4j -p 7474:7474 -p 7687:7687 \
  -e NEO4J_AUTH=neo4j/choose-a-password -e NEO4J_server_memory_pagecache_size=2G \
  -v icj_data:/data neo4j:5.26

docker exec icj-neo4j cypher-shell -u neo4j -p choose-a-password \
  "MATCH (c:Case) RETURN count(c);"   # 200 for icj-full (199 + stub case 201), 199 for icj-decisions

CALL db.index.vector.queryNodes('chunk_embedding', 10, $queryVector)   // query embedded with Qwen3-Embedding-8B
YIELD node, score RETURN node.id, node.text, score

Node counts for the decision graph (icj-decisions.dump):

Node	Count	Key relationships
Case	199	Document -IN_CASE-> Case
Document	2,407	Document -CITES_CASE-> Case (6,070 edges)
Chunk	121,438	Chunk -PART_OF-> Document; embedding property, vector-indexed
Provision	356	Document -CITES-> Provision (7,699 edges); Provision -PART_OF-> Instrument
Instrument	4	UN Charter, ICJ Statute, Rules of Court, Practice Directions
Judge	220	Document -SAT_ON-> Judge
Country	112	Document -APPLICANT / -RESPONDENT-> Country

icj-full.dump adds the pleadings and oral layers: Case 200, Document 6,396, Chunk 1,095,572. Citation edges are extracted from the decision layer only, so those counts match in both dumps. Citations were resolved to the right case by name and year, so same-named cases from different eras never cross-link.

The ten most-cited cases, over 6,070 case-to-case edges:

Case	Name	Cited
070	Military and Paramilitary Activities (Nicaragua v. USA)	292
091	Application of the Genocide Convention (Bosnia v. Serbia)	206
001	Corfu Channel (UK v. Albania)	162
008	Interpretation of Peace Treaties	149
064	US Diplomatic and Consular Staff in Tehran	138
104	LaGrand (Germany v. USA)	128
048	Northern Cameroons	127
016	Anglo-Iranian Oil Co.	115
050	Barcelona Traction	111
095	Legality of the Threat or Use of Nuclear Weapons	110

Provenance and license

• Historical decisions (1947 to Oct 2023): the CD-ICJ corpus by Sean Fobbe (Zenodo, DOI 10.5281/zenodo.3826444), CC0 1.0, 2,285 documents.
• Recent decisions (Oct 2023 to 2026) and all pleadings and oral records: retrieved from icj-cij.org.
• Case 2 (1947) is absent; its ICJ page has returned a server error for years.

Terms are per layer, not one blanket license. Read LICENSE before reuse; in short:

What	Terms
Decisions from CD-ICJ	CC0-1.0, passed through from upstream
Material retrieved from icj-cij.org (recent decisions, pleadings, oral)	No licence granted, since we hold none. Usage request: attribute the ICJ and this dataset, non-commercial, per the ICJ notice
The parsing, chunking, embeddings, graph, and tooling this project added	CC-BY-4.0

We claim no ownership of the underlying ICJ documents. The pleadings position is what we apply pending contact with the ICJ Registry, not a claim the material is licensed. Rights-holder requests will be honored.

Limitations

• is_ocr marks 119 wholly OCR-derived documents; ocr_page_count flags 464 more with isolated pages replaced. OCR coverage on damaged-but-inked pages is imperfect: some dense pages are transcribed incompletely.
• 51 documents are document_type: "unknown" (scans too degraded to classify); their text is still searchable.
• applicant/respondent in documents.jsonl are raw header strings; the graph's Country nodes are deduplicated and split into individual states. 52 Judge nodes carry a comma (a role suffix, or a jointly-signed opinion parsed as one name); neither affects the citation network.
• source_url is an internal provenance path, not a public link.

If you downloaded before 2026-07-29, re-download: an OCR pass had invented text on 942 blank pages across 37 pleadings files (now removed, decision layer and its citations were never affected), and is_ocr was corrected on 13 documents.

How it was built

Each PDF is rendered and read by an OCR model; a second pass classifies every line's role (title, paragraph, operative ruling, bench, signatures); a third extracts every citation of an earlier case and resolves it to a case id by name and year. Documents, passages, embeddings, and citation edges are written to Neo4j in one idempotent step.