Professional work frontier models still get wrong.
We're AI researchers working with practicing lawyers, clinicians, analysts and accountants. We build the expert data, evals and RL environments that teach models to get it right.
Talk directly to the researchers who build it.
Proud to partner with frontier AI teams
MiniMax
What we deliver
Off-the-shelf datasets
Ready-to-train RL environments for professional work. Each task ships with its environment, a weighted rubric, a reference answer and graded trajectories from frontier models, all signed off by a practicing expert.
Our legal dataset has 200+ tasks across 16 practice areas, with more produced to order.
- record/
- trial_transcript.pdf
- bodycam_log.xlsx
- briefs/
- petition_for_cert.pdf
- authorities/
- Mail 3
- Calendar
From: Dana Pike, Partner
Scope memo by Monday
Send me a 1,400-word scope memo on the qualified immunity question. Everything you need is in the record folder.
Due Mon 9:00
Each task ships with
- Environment
- Rubric
- Reference answer
- 18 graded runs
- Expert sign-off
Custom datasets and benchmarks
Tell us the domain, the workflow and the failures you care about. We build the environments, tasks and rubrics to your spec, calibrated against the models you want to beat, with a held-out benchmark that shows where they break.
| Criterion | Weight |
|---|---|
| Frames the controlling question | 12 |
| Cites the record accurately | 18 |
| Guard: doesn't repeat the “two warnings” claim | 24 |
| Meets length and format | 6 |
First-try pass rate on our legal dataset
Our expert network
Every task is reviewed and signed off by people who do this work for a living.
- LegalLitigation, appellate briefs, contracts, regulatory
- HealthcareClinical reasoning, documentation, coding and billing
- FinanceEquity research, modeling, credit, deal work
- AccountingAudit, tax, financial reporting
Tasks worth training on
- Hard
- Each task hides a plausible wrong path that the source files support. On our legal dataset, the best frontier model passes 18% of tasks on the first try, and in 52% of tasks none of 18 runs across six models passes.
- Solvable
- Every reference answer scores 100% on its own rubric, using only the files in the environment. The evidence that settles each trap is always inside it.
- Real
- A practicing attorney reviews and signs off on every legal task, spending about two hours on each. The facts are invented, so no client data is used.
- Diverse
- Failures span 11 distinct patterns, not one trick repeated. The legal dataset covers 16 practice areas, none above 12% of tasks.
How it works
Tell us what you need
The domain, the data format and roughly how much.
Get a sample
We send a sample within days so your team can judge it.
Scale to a full batch
Delivered in your format, with the grader and analysis for each task.