VISAI-AI/icj-precedent-cite-bench

ICJ Citation Prediction Benchmark

Predict which earlier cases, legal provisions and UN resolutions an ICJ decision cites from a description of the case. The test set contains 199 target decisions with reference citations and model inputs. Results include 8,955 answers from five models under three tool settings across three rounds.

The companion ICJ Citation Graph supplies the documents and citation links. Graph construction and the AI SDK sidecar are in the source repository.

Use the benchmark

pip install huggingface_hub
hf download VISAI-AI/icj-precedent-cite-bench --repo-type dataset --local-dir icj-bench
cd icj-bench
python score.py --pred results/predictions.jsonl --out scores.json

Scoring uses only the Python standard library and makes no API calls. Python 3.10 or later is required.

File	Contents
data/inputs.jsonl	Case descriptions and system/user messages to send to a model
data/test.jsonl	The same case descriptions with target identifiers, dates and reference citations
results/predictions.jsonl	All 8,955 completed answers with model, condition, round and validation metadata
results/metrics.json	Scores per round and their mean and sample standard deviation
protocol/	System prompt, tool lists and prediction schemas
manifest.json	Run identifiers, model names and source checksums

Use data/inputs.jsonl for inference. Reference citations and target decision identifiers in data/test.jsonl are for evaluation. Each input contains the case name, type, stage and year together with the parties' claims or the questions submitted for an advisory opinion.

To score your own answers, write one record per target in this format:

{"case_id":"001","target_document_id":"ICJ_001_CorfuChannel_GBR_ALB_1949-12-15_JUD_01_CO_00_EN","model_id":"my-model","condition":"my-setting","round":1,"prediction":{"citation_case_ids":[],"citation_provision_keys":[],"citation_resolution_symbols":[]}}

Copy the target identifier from data/test.jsonl. Citation values must be strings. Empty lists are valid answers. Each model/condition/round must cover all 199 targets with no duplicates. For a smaller diagnostic subset, pass --allow-partial; the report records how many targets are missing.

Evaluation

The reference lists are gold_citation_case_ids, gold_citation_provision_keys and gold_citation_resolution_symbols. Each family is scored separately by exact set matching. Numeric case identifiers are zero-padded to three digits. Provision keys and resolution symbols must match the reference spelling.

Precision, recall and F1 are computed from true positives, false positives and false negatives pooled across targets within each round. The summary reports the arithmetic mean and sample standard deviation across rounds. A zero denominator gives a score of zero. The scorer rejects missing citation fields so an incomplete record cannot silently become an empty answer.

The published results include 8,272 answers marked valid and 683 marked partial by the experiment parser. All contain predictions and are scored. Validation issues are preserved in the results; unmatched identifiers remain false positives.

Exact citation agreement measures how closely an answer reproduces the Court's references. A different citation may still be relevant, but these scores do not assess its legal suitability. Reference citations were extracted automatically and should not be treated as an exhaustive human-verified account of each decision.

Experimental setup

Condition	Tools available to the model
Parametric	None
Cypher + Semantic	Graph schema, read-only queries and passage search
Designed Tools Only	Twelve tools for looking up cases, documents, provisions, judges and UN resolutions

The exact tool names are in protocol/conditions.json. The reported runs used the Vercel AI SDK sidecar. Each target used a graph restricted to information available before its decision. Restoring the full companion graph alone does not reproduce that restriction; any new run must enforce the target's cutoff before exposing retrieval tools.

The five recorded model IDs are gpt-5.6-luna, deepseek-v4-flash-0731, deepseek-v4.1-flash, muse-spark-1.3-contributor and gpt-5-mini. The paper labels the Muse model muse-spark-1.3; the manifest preserves that mapping. Historical targets may have appeared in model pretraining, so temporal filtering of retrieval does not establish that a model has never seen a decision.

The included results and scorer reproduce the citation metrics without running inference. Full model/tool conversation traces and graph snapshots for individual cutoffs are not included in this dataset.

Reuse

Terms follow the source materials and the project's contributions as described in LICENSE. Run python -m unittest discover -s tests to check the scorer. File checksums are listed in SHA256SUMS.