Entropy Partners started as a simple conversation about renting a house. It turned into a focused attempt to fix the one thing that quietly breaks most legal AI systems: the data they learn from.
Most models are trained on case law archives and static statutes. Good for research, bad for live conversations. We work with practicing lawyers to create data that reflects real client questions, real edge cases, and the mistakes that should never make it into production. We do not just teach AI what is correct; we teach it what not to say.
Heads up: PDF translation is coming soon.
Legal chatbots don’t usually fail in obvious ways. They fail in subtle ones: fictional cases, mixed-up jurisdictions, confident answers that sound right until a lawyer looks closely. That happens because the training data was never designed for conversational use.
We don’t scrape court databases and call it training data. We talk to practicing attorneys about the questions clients actually ask and the mistakes they see repeatedly in real matters.
For every correct answer, we include carefully crafted wrong answers: the plausible hallucinations your model is likely to produce. This narrows the search space and cuts down confident-but-wrong outputs.
Need a specific jurisdiction, practice area, or language pair? We build with your lawyers and validate to your standards instead of forcing you into a generic off-the-shelf dataset.
We collaborate with seasoned attorneys to survey, author, refine, and annotate theoretical legal scenarios. Each pack includes grounded citations and structured labels so your model learns the why, not just the what. Which is tailor made to reduce hallucinations.
Tell us your jurisdiction(s), practice area, and target tasks/models.
We align on rights and price band before work begins.
We identify AI weak points with lawyers and create varied scenario Q&A with explicit correct/incorrect examples.
Issue, Rule, Application, Conclusion with custom labels as needed.
Double-checked by lawyers; a held-out test set includes at least one variation per law/case check.
JSONL or CSV to your spec. Minimum dataset: 5,000 rows.
Milestone-based with an upfront deposit.
Typical custom delivery ~3 months for the minimum scope.
Milestone-based delivery. Initial 15-20% deposit. Each acceptance unlocks the next deliverable.
Milestones and deliverables are agreed in the kickoff meeting. Larger programs extend in 2-week sprints.
Customers use our datasets to move from “interesting prototype” to systems their legal teams can actually trust.
Models trained on our data see substantial accuracy gains on real-world legal questions compared to generic web or case-law-only corpora.
Strategic wrong answers teach the model to recognize and avoid subtle legal mistakes instead of generating them confidently.
Attorney validation, citations, and metadata give you the documentation your compliance and legal teams demand.
Rows per Minimum Dataset
Jurisdictions Available Now
Structured Reasoning Labels
Imaginary Parties, Real Law
Available now: Tax Law (UAE, Jordan).
On request: Any country / additional areas (Contracts, Litigation, Regulatory, Compliance, Privacy, IP, Employment, M&A).
Minimum dataset: 5,000 rows.
Pricing: ranges from 5.98–100+ depending on jurisdiction, exclusivity, scope, and reviewer seniority.
(Exclusive categories & complex jurisdictions priced higher.)
Terms: milestone-based with an upfront deposit. Non-exclusive and exclusive licenses available (with terms that protect our IP and business).
Scenarios use imaginary parties; real law application is drafted and double-checked by lawyers.
Derived from our legal research and public information with citations and provenance.
Every order includes a held-out test set with variations per law/case check, reviewed by a lawyer.
Non-exclusive and exclusive packages with protective terms.
{"id":"uae-tax-000123",
"jurisdiction":"UAE",
"practice_area":"Tax",
"question":"A VAT-registered company receives an advance payment for a future supply. When is the output VAT due?",
"answer":"Under UAE VAT law, tax becomes due at the earlier of invoice issuance, receipt of payment, or supply date. Here, VAT is due upon receipt of the advance.",
"irac":{"issue":"VAT timing on advance","rule":"Tax due at earlier of invoice/payment/supply","application":"Payment received before invoice/supply -> VAT due now","conclusion":"Output VAT due upon advance"},
"citations":["UAE VAT Decree-Law No. 8 of 2017, Art. X"],
"labels":{"correct":true,"difficulty":"medium"},
"metadata":{"variant":"base","dataset_split":"train"}}
See how curated, IRAC-structured data reduces hallucinations and strengthens jurisdiction-specific reasoning.
IRAC scenarios with explicit correct/incorrect variants reduced hallucinations on tricky VAT timing prompts.
Adversarial variants improved robustness on LegalBench-style tasks across similar fact patterns.
Held-out test sets include at least one variation per law/case check, with lawyer verification.
| Feature | Ours | Generic Web Crawl |
|---|---|---|
| IRAC-Structured Q&A | Yes | No |
| Jurisdiction Tagging | Fine-grained | Sparse |
| Lawyer QA | Yes | No |
| Benchmark-Aware Design | Yes (LegalBench) | No |
The AI companies that win will not be the ones with the largest models, but the ones with the best data: data designed specifically for their domain, with deep practitioner input, clear provenance, and real-world mistakes baked in.
Every dataset we build compounds our advantage: our attorney network gets stronger, our pipelines get sharper, and your models benefit from a growing foundation of trusted legal understanding.
Train and test splits in JSONL or CSV. Scenario Q&A with IRAC fields (Issue, Rule, Application, Conclusion), citations, and negotiable metadata add-ons.
Pricing ranges from 5.98–100+ based on jurisdiction, exclusivity, scope, and reviewer seniority. Minimum dataset is 5,000 rows. Payments are milestone-based with a deposit.
UAE Tax and Jordanian Tax Law. We can produce custom datasets for any country on request.
We design examples to reflect patterns found in LegalBench-style tasks. See the paper and rankings.
No. Our datasets are for machine learning training and research. Scenarios use imaginary parties; the law application is drafted and reviewed by lawyers.
Reduce hallucinations and boost LegalBench-style performance with jurisdiction-specific, IRAC-structured data.
PDF translation is coming soon.