orgrctera/legalbenchrag_contractnli

LegalBenchRAG: ContractNLI (retrieval)

This dataset is the ContractNLI slice of LegalBench-RAG: a retrieval benchmark for legal RAG, built by tracing LegalBench-style queries back to minimal, human-annotated text spans in the underlying corpus. Each row poses a natural-language question grounded in a specific contract document; the target is to retrieve the exact character span(s) (and quoted text) that answer the question—matching the LegalBench-RAG evaluation philosophy (precise snippets rather than whole-document or coarse-chunk IDs).

Background: ContractNLI

ContractNLI (ContractNLI: A Dataset for Document-level Natural Language Inference for Contracts, Koreeda & Manning, Findings of EMNLP 2021) introduced document-level NLI for contracts: given a contract and a set of hypotheses (fixed statements about obligations, survival, confidentiality, etc.), a system must determine whether each hypothesis is Entailed, Contradicted, or Not mentioned by the contract, and (for non-neutral labels) identify evidence spans (sentences or list items) in the document.

The original release covers 607 non-disclosure agreements (NDAs) with 17 shared hypotheses annotated across documents. ContractNLI is notable for long, legally dense text, imbalanced labels, multi-task structure (NLI + evidence identification), and difficult linguistic phenomena such as negation by exception—characteristics that make both classification and retrieval challenging.

• Original resource: ContractNLI project page · GitHub: stanfordnlp/contract-nli
• License: CC BY 4.0 (see the official ContractNLI distribution for full terms).

Task: retrieval for ContractNLI

In this Hugging Face distribution, the task is retrieval aligned with LegalBench-RAG: for each input, a retriever should return the minimal relevant span(s) from the referenced contract—expressed here as character offsets into the source file together with the gold answer text and file path.

Downstream use cases include:

• Evaluating dense/sparse/hybrid retrievers on legal NDAs and fixed hypothesis templates.
• Training or tuning passage/span rankers where supervision is span-level, not only document-level.
• RAG pipelines where the generator must ground answers in precisely retrieved clauses (reducing context cost and hallucination risk, as argued in the LegalBench-RAG paper).

This is not the original three-way NLI classification task in isolation; it is the retrieval formulation used within the LegalBench-RAG benchmark family, with expected outputs that encode where in the contract the supporting language appears.

Relation to LegalBench and LegalBench-RAG

• LegalBench is a broad benchmark of tasks for measuring legal reasoning in language models; several tasks are derived from or inspired by ContractNLI-style settings.
• LegalBench-RAG (Pipitone & Alami, 2024) focuses on the retrieval step of legal RAG: 6,858 query–answer pairs over a large annotated corpus, emphasizing short, precise retrieved segments. Public artifacts are linked from the paper and community repositories (e.g. zeroentropy-ai/legalbenchrag).

When citing work built on this dataset, cite ContractNLI for the underlying annotations, LegalBench-RAG for the retrieval benchmark framing, and this dataset card for the specific HF packaging.

Dataset structure

Field	Description
id	Unique example identifier (UUID string).
input	Natural-language query, typically naming the agreement and asking a yes/no–style or obligation-focused question consistent with ContractNLI hypotheses (e.g. notice requirements, survival, confidentiality).
expected_output	JSON string encoding a list of objects. Each object includes a span [start, end] (character indices into the referenced file), the answer text for that span, and file_path locating the contract under the benchmark corpus layout (e.g. contractnli/...).

The default split contains 977 rows.

Example records

Example 1 (fields truncated for display; expected_output is a string containing JSON):

• input (excerpt):
  Consider the AGProjects' Non-Disclosure Agreement; Does the document require the Receiving Party to notify the Disclosing Party if they are required by law, regulation, or judicial process to disclose any Confidential Information?

• expected_output (parsed; one gold span):

  [
    {
      "span": [2895, 3428],
      "answer": "In the event that the RECEIVING PARTY is requested or required (by oral questions, interrogatories, requests for information or documents in legal proceedings, subpoena, civil investigative demand or other similar process) to disclose any of the CONFIDENTIAL INFORMATION, the RECEIVING PARTY shall provide the DISCLOSING PARTY with prompt written notice of any such request or requirement so that the DISCLOSING PARTY may seek a protective order or other appropriate remedy or waive compliance with the provisions of this Agreement. ",
      "file_path": "contractnli/AGProjects-NDA.txt"
    }
  ]
  

Example 2 — Schema illustration: each row follows the same pattern (id, full input string, expected_output JSON string with one or more {span, answer, file_path} objects). Evaluators typically parse expected_output and compare predicted spans or ranked passages against the gold spans.

References

1. ContractNLI (original dataset and NLI task)
   Yuta Koreeda and Christopher D. Manning. ContractNLI: A Dataset for Document-level Natural Language Inference for Contracts. Findings of EMNLP, 2021.
   ACL Anthology · PDF

   Abstract (short): The paper proposes document-level NLI for contracts: classify hypotheses as entailed, contradicted, or not mentioned, and identify evidence spans; it releases the largest annotated contract corpus to that point and introduces a strong span-based baseline (Span NLI BERT).

   @inproceedings{koreeda-manning-2021-contractnli-dataset,
     title = "{C}ontract{NLI}: A Dataset for Document-level Natural Language Inference for Contracts",
     author = {Koreeda, Yuta and Manning, Christopher},
     booktitle = {Findings of the Association for Computational Linguistics: EMNLP 2021},
     year = {2021},
     pages = {1907--1919},
     url = {https://aclanthology.org/2021.findings-emnlp.164}
   }
   

2. LegalBench-RAG (retrieval benchmark)
   Nicholas Pipitone and Ghita Houir Alami. LegalBench-RAG: A Benchmark for Retrieval-Augmented Generation in the Legal Domain. arXiv:2408.10343, 2024.
   arXiv · PDF

   Abstract (short): The paper argues for evaluating retrieval in legal RAG separately from generation, with minimal relevant segments annotated by legal experts so metrics can be computed precisely (including character-level alignment) and so LLMs can cite sources without huge imprecise chunks.

3. LegalBench (umbrella benchmark)
   See the LegalBench site and associated publications for the wider task suite.

Acknowledgments

Data lineage traces to ContractNLI (Stanford NLP / Hitachi America, Ltd. distribution terms) and the LegalBench-RAG retrieval construction. Use of the original ContractNLI materials remains subject to CC BY 4.0 and any additional terms stated on the official ContractNLI release.