Judicial Decision Modeling
Step-by-step reasoning traces that mirror how judges reach their decisions.
View details
Datasets for enterprises and AI labs
Expert Reasoning Data
We teach models to reason like domain experts. Our pipeline extracts step-by-step reasoning from expert documents in law, finance, medicine, and science, and specialists validate every trace. You get expert-quality training data at a lower cost than writing it from scratch.
- Datasets
- 12
- Domains
- 4
- Training methods
- SFT + RL
- Format
- JSONL
How it works
From expert documents to training-ready traces
Step 1
Extract
We gather expert documents such as rulings, filings, clinical cases, and transcripts, then use NLP to pull out the reasoning inside them.
Step 2
Structure
Each trace is organized into context, rules, facts, and a conclusion, so a model can learn every step.
Step 3
Validate
Domain experts review each trace before delivery. Starting from real expert work keeps costs well below writing data from scratch.
What we offer
Two ways to train expert reasoning
Supervised Fine-Tuning Datasets
Structured chain-of-thought traces that teach models to work through complex problems the way experts do.
- Training-ready reasoning traces in JSONL
- Domain datasets for legal, finance, medical, and science
- Every trace reviewed by a domain expert
Reinforcement Learning Environments
Interactive environments where AI agents learn domain reasoning through simulation and feedback.
- Environments for legal reasoning, diagnostics, and financial analysis
- Reward functions grounded in expert reasoning
- Built for training and evaluating agent decisions
Dataset catalog
Expert reasoning datasets by domain
Choose a domain and open any dataset to see what it contains. Datasets marked with the Hugging Face logo have a free public preview.
Legal, 3 datasets
Reasoning the way courts and regulators do
Traces from judicial rulings, procurement decisions, and constitutional review that teach models to apply rules to facts.
Legal dataset
Judicial Decision Modeling
What it captures
Court rulings broken into structured steps that map established facts to statutes, apply precedent, and handle exceptions explicitly.
What the model learns
To apply the law to the facts and weigh precedent systematically, instead of producing text that only sounds legal.
Who uses it
Law firms building specialized models that predict case outcomes more accurately than general frontier models.
See a worked example on US appellate law- Format
- JSONL
- License
- Proprietary enterprise
Federal Government Bid Protests (US)
Logical audits of federal contract awards against procurement regulations.
View detailsFree preview
Free previewLegal dataset
Federal Government Bid Protests (US)
What it captures
GAO bid protest decisions turned into deductive chains that match each vendor claim to the Federal Acquisition Regulation (FAR) evaluation criteria.
What the model learns
To check technical scoring and agency compliance precisely, without the guesswork typical of general models.
Who uses it
GovCon vendors and defense contractors auditing proposals for FAR compliance and assessing whether a protest is worth filing.
- Format
- JSONL
- License
- Proprietary enterprise
Constitutional Compliance (FR)
French civil-law reasoning on the constitutionality of proposed legislation.
View details
Legal dataset
Constitutional Compliance (FR)
What it captures
Articles of a projet de loi tested against the French Constitution and the bloc de constitutionnalité, flagging overreach and conflicts with fundamental rights.
What the model learns
Native French civil-law reasoning, written in French, without importing US legal concepts.
Who uses it
Public bodies, think tanks, and European compliance teams screening bills for constitutional risk before they are enacted.
- Format
- JSONL
- License
- Proprietary enterprise
Finance, 3 datasets
From market signals to defensible conclusions
Traces from analyst research, executive disclosures, and enforcement cases that teach models to reason about value and risk.
Equity Investment Thesis
Market data, valuation multiples, and macro trends synthesized into buy, hold, or sell rationale.
View details
Finance dataset
Equity Investment Thesis
What it captures
How analysts weigh quantitative metrics against qualitative factors to build a coherent investment argument.
What the model learns
To reason through conflicting market signals without basic logical errors.
Who uses it
Hedge funds and asset managers generating structured investment memos across large equity universes.
- Format
- JSONL
- License
- Proprietary enterprise
Executive Rationale Extraction
The business reasoning behind management's forward-looking financial projections.
View details
Finance dataset
Executive Rationale Extraction
What it captures
Management's narrative on operating trends, liquidity, and capital resources, linked to its forecasts across real public companies.
What the model learns
To test forward-looking statements against operating data and isolate the real drivers of a company's performance.
Who uses it
Financial institutions parsing annual reports at scale to surface core business rationale and risks that general models miss.
- Format
- JSONL
- License
- Proprietary enterprise
Regulatory Fraud Detection
Financial irregularities traced to the securities regulations they violate.
View details
Finance dataset
Regulatory Fraud Detection
What it captures
Accounting anomalies, insider trading signals, and disclosure gaps, each linked to the SEC rule it breaches.
What the model learns
To connect small financial discrepancies to material legal exposure.
Who uses it
Audit firms and compliance teams flagging potential violations in internal communications and trading logs before regulators do.
- Format
- JSONL
- License
- Proprietary enterprise
Medical, 3 datasets
Clinical reasoning you can audit
Traces from diagnostic cases, treatment decisions, and regulatory reviews that show every step between evidence and decision.
Differential Diagnostics & Synthesis
Clinical reasoning that combines symptoms, imaging, labs, and genomics into a diagnosis and outcome.
View detailsFree preview
Medical dataset
Differential Diagnostics & Synthesis
What it captures
How specialists cross-reference clinical presentation with radiology, pathology, and genetic findings to rule out competing hypotheses.
What the model learns
To produce an auditable diagnostic rationale that carries through to treatment and long-term outcomes.
Who uses it
Healthcare organizations and medical AI teams building clinical decision support grounded in traceable evidence.
- Format
- JSONL
- License
- Proprietary enterprise
Clinical Therapeutics
Symptoms, contraindications, and patient history mapped to the right drug treatment.
View details
Medical dataset
Clinical Therapeutics
What it captures
Prescribing decisions that weigh presenting symptoms against allergies and current medications to select the safest effective option.
What the model learns
To justify each recommendation step by step and put contraindications ahead of common but unsafe drug associations.
Who uses it
Healthcare enterprises deploying decision support that returns auditable, protocol-aligned treatment recommendations.
- Format
- JSONL
- License
- Proprietary enterprise
FDA Medical Clearance
Device and clinical trial data evaluated against FDA safety and efficacy standards.
View details
Medical dataset
FDA Medical Clearance
What it captures
Device specifications, risk classification, and trial results matched to the FDA requirements that apply to them.
What the model learns
To spot missing safety data, assess trial endpoints, and anticipate regulatory pushback.
Who uses it
Biotech and medical device companies preparing stronger submissions in less time.
- Format
- JSONL
- License
- Proprietary enterprise
Science, 3 datasets
Rigor for research, safety, and invention
Traces from peer review, incident investigations, and patent examination that teach models to test claims and trace causes.
Academic Peer Review
Methodological critique and validity assessment of research papers.
View details
Science dataset
Academic Peer Review
What it captures
Validated reviewer reasoning that leads to an accept or reject decision with a clear rationale.
What the model learns
To critique research rigor rather than restate the abstract.
Who uses it
Publishers and R&D teams screening papers to separate sound findings from work that needs revision.
- Format
- JSONL
- License
- Proprietary enterprise
Incident Root-Cause Analysis
Multi-step causal analysis of mechanical, human, and environmental factors from NTSB reports.
View detailsFree preview
Science dataset
Incident Root-Cause Analysis
What it captures
How investigators link environmental conditions, mechanical tolerances, and operator decisions to a single root cause.
What the model learns
To separate proximate causes from systemic flaws without jumping to conclusions.
Who uses it
Insurance, supply chain, and safety engineering teams assessing incident reports, liability, and recurring risks.
- Format
- JSONL
- License
- Proprietary enterprise
Patent Prosecution Rationale
Why patent examiners accept or reject claims, based on statutory requirements.
View details
Science dataset
Patent Prosecution Rationale
What it captures
USPTO examiner reasoning that tests claim language against novelty, utility, non-obviousness, and prior art.
What the model learns
To read technical claims like an examiner and flag overbroad language or prior-art overlap.
Who uses it
IP firms and corporate R&D teams pre-screening applications before filing to avoid costly rounds with the USPTO.
- Format
- JSONL
- License
- Proprietary enterprise
Custom Dataset Extraction
Need a domain we don't cover yet? We build custom reasoning datasets with research labs and enterprise teams: scoped to your use case, extracted from your documents, and validated by domain experts before delivery.
Get in touchFine-tuning on Legal Reasoning Traces
How a model fine-tuned on our legal traces outperforms leading general models at predicting complex case outcomes.
Read the blog post