HumanEdgeAI/Corp_FinanceLongContextReasoning

Corp_Finance Long-Context Reasoning — Expert-Authored Credit Agreement Benchmark (Showcase Sample)

A five-record public sample from Corp_Finance Long-Context Reasoning, a subject-matter-expert benchmarking dataset built by Human Edge for evaluating frontier model reasoning over leveraged finance and syndicated credit documentation.

Every question, reasoning trace, and answer in this dataset was authored by a practicing finance professional and independently reviewed by 2-3 others. Nothing here is model-generated.

This is a curated showcase, not a training corpus. It exists to make the authoring and review methodology inspectable. See Contents.

• Curated by: Human Edge
• Language: English
• License: CC BY 4.0
• Repository: HumanEdgeAI/Corp_FinanceLongContextReasoning

---

Why this dataset exists

Frontier models are fluent on credit documentation and unreliable on it. They resolve defined terms to generic market meanings instead of the agreement's own, stop at the first hop of a definition chain, drop qualifying conditions when summarizing, and make arithmetic errors on covenant calculations. These failures are hard to detect precisely because the output reads well.

The difficulty is structural. A credit agreement runs to hundreds of pages, and the information needed to answer a single question is rarely in one place: a covenant threshold in one section depends on a defined term in another, which carves out an exception defined in a third. Answering correctly means holding the whole chain at once — which is why this dataset is built around long-context reasoning rather than retrieval.

Catching those failures requires evaluators who already know what the right answer looks like. This dataset is built around that constraint: expert authorship, expert review, and a reasoning trace that has to be defensible step by step against a cited section.

---

Contents

Records in this sample	5
Records in the full corpus	33
Source documents	2
Fields per record	11
Format	Parquet (Snappy)
Split	train (single split; no train/test partition — this is an evaluation set)

The five records were selected to (a) maximize reviewer-assigned Overall Quality and (b) span the widest possible range of question types and both source documents. Where a task identifier existed in both documents, the higher-scoring version was taken.

Records included

Question ID	Document	Question type	Difficulty (R1 / R2)	Overall Quality (R1 / R2)
Task #3 - Q&A #7	Doc #1 — Leveraged Finance	Summarization	4 / 5	5 / 5
Task #3 - Q&A #2	Doc #1 — Leveraged Finance	Market Opinion	4 / 5	5 / 5
Task #2 - Q&A #6	Doc #2 — Syndicated Credit	Industry Jargon	5 / 4	5 / 4
Task #1 - Q&A #2	Doc #1 — Leveraged Finance	Numeric Reasoning	4 / 4	4 / 4
Task #3 - Q&A #6	Doc #2 — Syndicated Credit	Multi-Section Reference	3 / 4	4 / 4

Mean Overall Quality across the ten independent reviewer ratings in this sample: 4.5 / 5. All five records passed all four review dimensions under both reviews; this file carries the citation-correctness verdicts, with the other three dimensions omitted as uniformly Pass across the sample.

Question Type is a field of the full corpus but is not a column in this sample file. The types above are stated here for orientation.

---

Source documents

Both source agreements are real, publicly filed credit agreements, available in full from the SEC's EDGAR system. No part of this benchmark depends on a private document.

	Doc #1	Doc #2
Instrument type	Leveraged Finance	Syndicated Credit
Borrower	Novelis Holdings Inc.	Seagate HDD Cayman
Parent	Novelis Inc.	Seagate Technology Holdings plc
Administrative Agent	Citibank, N.A.	The Bank of Nova Scotia
Dated	11 March 2025	30 January 2025
Facility	$1.25bn Term Loan B	$1.3bn revolving credit facility
Filing	EX-10.9 to Novelis Inc. Form 10-K, filed 12 May 2025	EX-10.1 to Seagate Technology Holdings plc Form 8-K, filed 3 February 2025
Full text	SEC EDGAR	SEC EDGAR

The two documents were chosen to contrast: a secured leveraged term loan with an extensive covenant package and multi-jurisdiction guarantee structure, against a large investment-grade-style syndicated revolver whose collateral obligations are themselves rating-triggered. Both run to several hundred pages.

Every record is independently verifiable. Because the agreements are public, any reader can open the cited section and check the reasoning trace against the actual contractual language — the claim and its evidence are both inspectable. Doc #1 records cite sections including 5.15 (Designation of Subsidiaries), 6.04 (Investments, Loans and Advances) and 6.07 (Dividends); Doc #2 records cite sections including 1.01 (Defined Terms) and 2.17 (Defaulting Lenders). Note that section numbering differs between the two agreements, so a section reference is only meaningful alongside its Document ID.

Human Edge built this benchmark on public filings by design, so that the work can be shared, audited, and reproduced without an NDA. The full corpus draws on these same two agreements.

---

Question type taxonomy

The full corpus is stratified across six categories, each targeting a distinct, documented model failure mode. Five of the six are represented in this sample.

Summarization — synthesis across multiple clauses into an accurate summary. Targets the tendency to omit qualifying conditions or conflate related-but-distinct provisions. Requires genuine synthesis, not verbatim reproduction of one paragraph.

Multi-Section Reference — tracing a definition chain across two or more sections. Targets failure at the second or third hop: stopping at the first definition found, or hallucinating the contents of the next section. Chains must not be short-circuitable.

Industry Jargon — resolving a defined term to this agreement's contractual meaning and identifying how it departs from standard market usage. Targets the default to generic definitions without consulting the document.

Numeric Reasoning — multi-step financial calculation using values drawn from the document. Targets arithmetic errors, skipped steps, and values pulled from the wrong section.

Market Opinion — combining document evidence with professional judgment to evaluate a provision against market norms. Targets the two opposite evasions: restating the document without evaluating it, or asserting a market view without grounding it in the text.

Basic Extraction (not in this sample) — locating an explicitly stated fact. A deliberate floor-coverage baseline against which higher-complexity performance is measured. Citation and reasoning requirements are identical to every other type.

---

Who wrote and reviewed these tasks

Every question, reasoning trace, answer, and review rating in this dataset was produced by a practicing corporate finance professional. Six subject-matter experts contributed.

The cohort was recruited against a defined profile rather than a general finance background:

Domain	Leveraged finance and syndicated credits
Role	Credit analyst or financial analyst
Experience	3–8 years at a bulge-bracket bank, credit fund, or law firm
Geography	United States

The experience band is deliberate. These questions test whether a model can navigate a 200-page credit agreement the way a working analyst does — locating a definition, tracing it through cross-references, and computing a result. That is the daily work of an analyst at this level, and specifying the band narrowly produced a cohort calibrated to the task rather than to seniority.

Before receiving assignments, every expert completed a three-module training program covering how language models work, the expert's role in evaluation, and the question type taxonomy used in this dataset. Training performance was tracked and had to clear a threshold for an expert to be assigned work.

Each expert performed two distinct roles on separate documents: authoring — reading a full agreement and writing eight question-and-answer pairs with complete reasoning and clause-level citations — and peer review — independently scoring pairs written by other experts. No expert reviewed their own work.

---

Quality assurance

Authoring, machine pre-screening, and human review are separate stages performed by different parties. No stage grades its own work, and no rating published in this file was produced by the person who wrote the record.

Stage	Performed by	Produces
1. Authoring	Task Author (finance SME)	Question, Reasoning Steps, Answer
2. Automated pre-screen	Binary: Type tag validation, Duplicate detection, Answerability screen, Citation correctness, Answer completeness, Factual accuracy	Scored: Clarity and reasoning-completeness gating
3. Independent review ×2–3	Task Reviewers (finance SMEs)	Four P/F review dimensions (of which Citation Correctness R1/R2 is included here) plus two scored criteria, Overall Quality (R1/R2) and Difficulty (R1/R2), both included here

Reviewers work independently and are not shown each other's verdicts, which is why both ratings are published separately rather than averaged.

Records that fail review are revised and re-reviewed before inclusion in the corpus. Every record in this sample cleared all four dimensions under at least two independent reviews.

---

Field reference

Identifiers

Field	Type	Description
Question ID	string	Task and Q&A index, e.g. Task #3 - Q&A #7. Unique only in combination with Document ID — the same identifier appears against both source documents in the full corpus.
Document ID	string	Source agreement: Doc #1 (Leveraged Finance) or Doc #2 (Syndicated Credit).

Content

Field	Type	Populated by	Description
Question	string	Task Author	The stimulus posed to the model.
Reasoning Steps	string	Task Author	The logical path, intermediate computations, and document lookups leading to the answer. Each step pairs a step description with the citation supporting it, flattened to numbered readable text in this file.
Answer	string	Task Author	The response that satisfies the question.

Reviewer ratings

R1 and R2 denote the first and second independent reviewer throughout this dataset. All four columns are populated by Task Reviewers.

Field	Type	Description
Difficulty (R1)	int32	First reviewer's complexity rating, 1–5.
Difficulty (R2)	int32	Second reviewer's complexity rating, 1–5.
Overall Quality (R1)	int32	First reviewer's overall quality rating, 1–5.
Overall Quality (R2)	int32	Second reviewer's overall quality rating, 1–5.

Difficulty scale:

1	Direct extraction from a single, easily located clause. No interpretation.
2	Explicit information requiring careful reading, or located in a less-prominent section.
3	Synthesis across multiple sub-clauses, simple calculation, or basic cross-referencing.
4	Multi-step reasoning, integration across separate sections, or professional judgment against market standards.
5	Complex chains (4+ hops), advanced calculations with term resolution, or nuanced market evaluation requiring expert interpretation.

Review verdicts

Two columns, both Pass / Fail, populated by Task Reviewers.

Field	Description
Citation Correctness P/F (R1)	First reviewer: do the cited sections genuinely support the answer?
Citation Correctness P/F (R2)	Second reviewer: same question, assessed independently.

Citation correctness is scored separately from factual accuracy by design. An answer can be entirely correct while pointing at the wrong section, and that combination is the failure most likely to survive casual review — which is exactly why it is the dimension retained here.

Review covers four dimensions in total. The other three — factual accuracy, answer completeness, and reasoning trajectory — are recorded in the full corpus and are Pass for every record in this sample under both reviews.

Personal and sensitive information

The dataset contains no private personal data. It is built entirely on credit agreements filed publicly with the SEC, which name corporate entities — borrowers, parents, administrative agents — and, in places, individuals acting in a professional capacity such as counsel or signatories. Any such name is already a matter of public record in the filing itself; nothing here discloses information that was not already public. Author and reviewer identities are not published.

Intended uses

Appropriate for: inspecting expert-authored evaluation methodology; understanding how reasoning traces and citation requirements are specified for financial documentation; assessing whether this annotation standard fits an evaluation program.

Not appropriate for: benchmarking or scoring a model. Five records support no statistical claim whatsoever. Fine-tuning.

---

Limitations

• Sample size is 5 (of 33 in total). It is illustrative by construction.

• Selection is deliberately biased toward high-scoring records. The full corpus contains records that failed review dimensions on first pass; this sample contains none.

• Two source documents, both US-market credit agreements. No coverage of other instrument types, jurisdictions, or governing law.

• Basic Extraction is absent, so the difficulty range is skewed upward relative to the full corpus.

• All content is English.

• Some corpus fields are withheld from this sample, to keep it readable and to publish only independently-assigned ratings — not because of any concern about the values:

  • Question Type — the taxonomy label (stated above per record instead)
  • Author self-assessed difficulty — the task author's own complexity rating of their own work. Excluded so that every rating published here is an independent one.
  • Aggregate quality score — the mean of the two reviewer quality ratings. Excluded in favor of publishing both underlying ratings, which carry strictly more information.
  • Factual Accuracy P/F, Answer Completeness P/F, Reasoning Trajectory P/F (R1 and R2) — three of the four review dimensions, uniformly Pass across these five records
  • Automated pre-screen layer — clarity and reasoning-completeness scores with their written justifications, plus the overall pass flag
  • Pre-revision submissions — the originally submitted question, reasoning steps, and answer, which allow before/after inspection of what review actually changed
  • Pipeline bookkeeping — review counts and inclusion flags

  The full corpus, including the pre-revision fields, is available under NDA.

---

License

Released under the Creative Commons Attribution 4.0 International license (CC BY 4.0).

You are free to share and adapt this dataset for any purpose, including commercially, provided you give appropriate credit to Human Edge, link to the license, and indicate whether changes were made.

The license covers Human Edge's contribution — the questions, reasoning traces, answers, and review ratings. It does not extend to the underlying credit agreements themselves, which are the parties' own documents and are reproduced here only by reference to their public SEC filings. It does not extend to the remainder of the corpus.

Citation

If you use this dataset, please cite it:

BibTeX:

@misc{humanedgeai2026corpfinance,
  title     = {Corp\_Finance Long-Context Reasoning: Expert-Authored Credit Agreement Benchmark (Showcase Sample)},
  author    = {{Human Edge}},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/HumanEdgeAI/Corp_FinanceLongContextReasoning}
}

APA:

Human Edge. (2026). Corp_Finance Long-Context Reasoning: Expert-Authored Credit Agreement Benchmark (Showcase Sample) [Data set]. Hugging Face. https://huggingface.co/datasets/HumanEdgeAI/Corp_FinanceLongContextReasoning

About Human Edge

Human Edge builds expert human data for AI development — SME-based evaluation, benchmarking, and reinforcement learning from expert feedback in domains where correctness requires professional judgment: finance, legal, healthcare, and tax.

This dataset is a sample of the pilot phase of a larger program. The production methodology scales the cohort, the review pipeline, and the volume well past what is shown here.

To discuss an evaluation or benchmarking engagement: humanedgetech.ai