HumanEdgeAI/LegalReasoning

Human Edge — Legal Reasoning Evaluation (Showcase Sample)

A public 6-task sample from a rubric-based legal reasoning evaluation dataset built by Human Edge (humanedgetech.ai). Each task is authored and reviewed by practicing senior lawyers and is designed to produce a verifiable, per-criterion reward signal for post-training and evaluation of frontier language models on high-complexity legal work.

The sample contains one task per legal subdomain, drawn from a larger internal corpus of 50 expert-authored tasks.

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

Why this dataset exists

Most legal evaluations reduce to multiple-choice recall or bar-exam-style questions. Real legal work does not look like that: it is open-ended, multi-jurisdictional, and judged on reasoning quality, appropriate hedging, and whether the sources actually say what the answer claims they say.

This dataset takes the opposite approach. Every task is a realistic instruction from a practicing lawyer's desk, paired with an expert-authored reference answer and a fine-grained rubric of weighted, binary criteria totalling exactly 100 points per task (29–43 criteria in this sample). The rubric — not the reference answer — is the object of evaluation. That makes scoring reproducible across graders and directly convertible into per-dimension reward signals.

Task architecture

Each row is a triple-component unit:

Component	Field	What it is
1. Legal reasoning prompt	prompt	A realistic scenario requiring multi-step reasoning and professional judgment, with role, jurisdiction, and deliverable format specified
2. Golden answer	golden_answer	An expert-authored reference analysis (17k–38k characters in this sample)
3. Scoring rubric	rubrics	29–43 categorized, weighted, binary criteria per task, positive weights totalling 100 points, each with the author's justification

Rubric design

Criteria are binary and weighted. Positive weights reward required content; negative weights penalize specific failure modes (hallucinated authority, overconfident advice, missing caveats).

Every task's rubric is calibrated to the same point budget. Positive criteria total exactly 100 points, allocated across the three positive categories in a fixed split:

Category	Sub-category	Point budget	Criteria	What it tests
Substance		65	108
	Explicit requirements		27	Did the answer do what was literally asked?
	Implicit requirements		27	Did it surface what a competent practitioner would raise unprompted?
	Legal correctness		22	Is the law stated accurately?
	Reasoning quality		32	Is the analytical path sound, not just the conclusion?
Sources & References	—	20	34	Are cited authorities real, relevant, and correctly characterized?
Structure & Style	—	15	32	Does it read like professional work product?
Negative Criteria	—	penalties	25	Penalties for specific, anticipated failure modes
Total		100	199

The point budget holds for all six tasks individually, not just in aggregate — so a model's raw score is already a percentage, and scores are directly comparable across tasks and subdomains despite differing criteria counts. Negative criteria sit outside the 100 points and subtract from the earned total, so a score can fall below zero.

The criteria counts vary by task (29–43) because experts allocated the fixed budget at whatever granularity the material demanded — a task needing many small checks uses more +1 criteria, one turning on a few decisive points uses +5s.

The rubric_category field stores each of these as a single literal string. Filter on these values exactly:

Substance Rubric - Explicit requirements, Substance Rubric - Implicit requirements, Substance Rubric - Legal correctness, Substance Rubric - Reasoning quality, Sources and References Rubric, Structure & Style Rubric, Negative Rubric

Weight distribution across the 199 criteria in this sample: +5 (71), +3 (71), +1 (32), -1 (2), -3 (7), -5 (16).

Contents

Six tasks, one per subdomain:

task_id	area	Rubric criteria	Peer review overall quality (R1 / R2)
18	Employment & Labor	34	3 / 4
40	Intellectual Property (IP)	43	4 / 4
68	Commercial Litigation	34	4 / 5
82	Regulatory & Compliance	29	4 / 5
93	Corporate & M&A	30	4 / 4
100	Contract Law	29	4 / 5

Jurisdictions are US and UK (England and Wales). task_id values are the original corpus identifiers and are intentionally non-contiguous.

Who wrote and reviewed these tasks

Tasks were authored by senior legal practitioners recruited against a hard credential bar: 8+ years in practice, a Master's or PhD in law, and a background in AmLaw 100 / Magic Circle firms, senior courts, government, or in-house at large enterprises.

The contributing cohort averaged ~18.5 years of practice, split roughly 75% US / 25% UK, with education spanning T14 US law schools and leading UK and European universities. Identity was verified for every contributor before any project interaction.

Domain expertise alone does not make a calibrated evaluator, so every contributor completed a mandatory training program (~2.3 hours average) on stress-testing model outputs, applying rubrics consistently, and documenting rationale.

Quality assurance

Three stages, applied to every task in the source corpus:

1. Automated checks. Every task passes programmatic validation before it reaches a reviewer. Prompts are checked for scenario framing, role specification, deliverable format, and jurisdictional context. Rubrics are checked for binary format compliance and a minimum criterion count. A model-graded pass confirms each rubric is consistent with its golden answer.
2. Double-blind expert peer review. Two independent senior practitioners review each task, rating overall quality, difficulty, representativeness, and task specification quality, each with written rationale. Reviews drive iterative revision rather than a simple accept/reject: the author revises against reviewer comments, and the version shipped here reflects those revisions.
3. Difficulty validation. Tasks are stress-tested against frontier models. A task is admitted only if strong models still fail a meaningful share of its rubric criteria — tasks that models solve comfortably carry no training signal and are rejected.

Tasks in the source corpus averaged roughly eleven hours of expert effort each, counting authoring, peer review, and revision. Reviewer ratings clustered tightly, indicating consensus among practitioners rather than averaged-out disagreement.

Field reference

Task content

Field	Type	Description
task_id	int32	Original corpus identifier
area	string	Legal subdomain
prompt	string	The legal reasoning prompt given to the model
golden_answer	string	Expert-authored reference answer
associated_rubrics	int32	Number of rubric criteria; always equals len(rubrics)
rubrics	list<struct>	Nested rubric criteria (see below)
task_details	string	Author's note on why the task is hard or interesting
reference_materials	string	Source authorities relied on; free-text, sometimes URLs, sometimes citations

rubrics struct

Field	Type	Description
rubric_category	string	One of the seven literal values listed above
score_option	int32	Weight: +5, +3, +1, -1, -3, -5
criterion	string	Binary question applied to the model's answer
justification	string	Author's rationale for why this criterion matters; empty for 5 of 199 criteria (see Limitations)

Text has been normalized: non-breaking spaces, stray tabs and other exotic whitespace introduced by the authoring tools have been folded to plain spaces. Newlines in golden_answer are preserved.

Author commentary

The task author's own notes on realism and difficulty. Their numeric self-ratings are deliberately not included — an author's rating of their own task is not independent evidence, so the only ratings in this dataset come from the two peer reviewers.

Field	Type	Description
realistic_explanation	string	Why the scenario reflects real practice
difficulty_explanation	string	What makes the task hard, and where models are expected to fail

Peer review

Two independent reviewers per task, both prefixed: peer_review_1_ is the first-round review (conducted on the first draft version), peer_review_2_ the second (conducted on the second draft version). The final version shipped here was produced after revising comments from the second reviewer. Reviewer identities are not published.

The *_label fields carry the rating exactly as the reviewer selected it — score and wording together, e.g. 4 - Above Standards. To get a numeric value, split on the first -.

Field suffix	Type	Description
overall_quality_label	string	Rating 1–5, e.g. 4 - Above Standards
overall_quality_description	string	Reviewer's written rationale for the quality rating
representativeness_label	string	Rating 1–5 — how typical this task is of real practice, e.g. 5 - Core Activity
difficulty_label	string	Rating 1–5, e.g. 3 - Moderate
difficulty_description	string	Reviewer's written rationale for the difficulty rating
task_specification_quality_label	string	Categorical: Well-Specified, Overspecified, or Underspecified

Rating scales, as presented to reviewers:

• Overall quality: 1 - Far Below Standards, 2 - Below Standards, 3 - Meets Minimum Standards, 4 - Above Standards, 5 - Exemplary
• Representativeness: 1 - Rarely Encountered, 2 - Uncommon, 3 - Somewhat Typical, 4 - Typical, 5 - Core Activity
• Difficulty: 1 - Trivial → 5 - Very Hard (3 - Moderate, 4 - Hard)
• Task specification quality: Underspecified, Well-Specified, Overspecified

Personal and sensitive information

The dataset contains no personal data. Scenarios are built on fictional parties and hypothetical facts; the only real names are the litigants in cited case law, which is public record. Author and reviewer identities are not published.

Intended uses

• Evaluating LLM performance on open-ended, high-complexity legal reasoning
• Rubric-as-reward research: RLVR, reward model training, process supervision
• LLM-as-judge calibration — the rubrics give a judge concrete, verifiable criteria instead of a vague quality prompt
• Studying legal-domain failure modes, especially source and citation reliability
• A reference template for anyone constructing expert-authored rubric evaluations in other professional domains

Out of scope: legal advice of any kind. The golden answers are evaluation artifacts written against hypothetical facts, not guidance on any real matter, and must not be relied on as such.

Limitations

• Six tasks. This is a methodology sample, not a benchmark. Do not report aggregate scores over six tasks as a model capability claim.
• Scoring requires a judge. Criteria are binary but not mechanically checkable; they need a competent grader, human or model. Judge choice will shift absolute scores.
• US and UK only. Nothing here generalizes to civil law systems, EU-level practice, or other common law jurisdictions without revalidation.
• Contamination risk. Published openly, these tasks may enter future training corpora. Treat results on this sample as indicative once models trained after its publication are involved.
• Reviewer identities are withheld. Cohort-level credentials are described above; individual names and affiliations are not published.
• Five criteria carry no justification. Four in task 40 and one in task 100 have an empty justification; the authors did not record one. The criteria themselves are complete, scored, and counted in the point budget — only the explanatory note is absent. Every other field is populated in every row.

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 prompts, golden answers, rubrics, and review ratings. It does not grant rights in the third-party statutes, regulations, and court opinions cited within the tasks; those remain governed by their own terms. It does not extend to the remainder of the corpus.

Citation

If you use this dataset, please cite it:

BibTeX:

@misc{humanedgeai2026legalreasoning,
  title     = {Human Edge Legal Reasoning Evaluation: Showcase Sample},
  author    = {{Human Edge}},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/HumanEdgeAI/LegalReasoning}
}

APA:

Human Edge. (2026). Human Edge Legal Reasoning Evaluation: Showcase Sample [Data set]. Hugging Face. https://huggingface.co/datasets/HumanEdgeAI/LegalReasoning

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