moriyad's picture
Initial release
089f7a6 verified
|
Raw History Blame Contribute Delete
3.55 kB
metadata
license: cc-by-4.0
language:
  - en
pretty_name: Contract Clause Graphs (GRAPH-GRPO-LEX)
size_categories:
  - 1K<n<10K
task_categories:
  - text-generation
tags:
  - legal
  - contracts
  - knowledge-graph
  - information-extraction
  - cuad
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train.parquet
      - split: validation
        path: data/validation.parquet
      - split: test
        path: data/test.parquet

Contract Clause Graphs

Each row pairs one contract clause with a graph of the clauses, parties, defined terms and values it mentions. This is the training and evaluation data of GRAPH-GRPO-LEX (IEEE ICDMW 2025).

Paper · arXiv · Code and contract linter

Split Clauses Contracts
train 2,071 35
validation 262 3
test 266 4

Splits are by contract: no contract and no clause text appears in more than one split.

Use

import json
from datasets import load_dataset

ds = load_dataset("moriyad/contract-clause-graphs")
row = ds["train"][0]
graph = json.loads(row["graph"])     # {"contract_id": ..., "nodes": [...], "edges": [...]}

Fields

Field Description
id, contract_id, clause_id Row, contract (CUAD file name) and clause identifiers
clause_title, clause_text, clause_chars The clause and its length in characters
graph The graph, as a JSON string
n_nodes, n_edges Graph size
in_standard, in_grpo Subsets used in the paper for supervised fine-tuning (2,385 rows, graphs up to 15 nodes) and for GRPO (1,640 rows, up to 9 nodes)
in_contractgrapheval The 150 validation rows used in the ContractGraphEval benchmark
is_fragment Clause text under 20 characters, such as a bare clause number (88 rows, 3.4%)

Nodes: CLAUSE, PARTY, DEFINED_TERM, VALUE. Edges, each from a clause: IS_PART_OF and REFERENCES (to a clause), DEFINES and USES (to a defined term), MENTIONS_PARTY (to a party), CONTAINS (to a value).

The labelling prompt is in instruction.txt.

How it was made

  • Source: 42 commercial agreements from CUAD, split into clauses automatically.
  • Labels: generated by GPT-5 (gpt-5-chat-latest) from the prompt above.
  • Validation: the alternative annotator test (Calderon et al., 2025) with 3 human annotators on 43 clauses. The model agreed with the annotators' consensus about as well as a third annotator did (ω = 1.0 at q = 0.05, ε = 0.15; ρ = 0.91).

Limitations

  • The graphs are model-generated and were validated on a sample, not checked one by one.
  • 42 English-language commercial contracts from U.S. securities filings.
  • Each graph covers a single clause.

Licence and citation

cc-by-4.0. The contracts come from CUAD (The Atticus Project, CC BY 4.0).

@inproceedings{dechtiar2025graphgrpolex,
  title     = {GRAPH-GRPO-LEX: Contract Graph Modeling and Reinforcement Learning with Group Relative Policy Optimization},
  author    = {Dechtiar, Moriya and Katz, Daniel Martin and Sundaresan, Mari and Jaume, Sylvain and Wang, Hongming},
  booktitle = {2025 IEEE International Conference on Data Mining Workshops (ICDMW)},
  pages     = {767--776},
  year      = {2025},
  doi       = {10.1109/ICDMW69685.2025.00092}
}