01 Source
Source the knowledge. Vetted experts and enterprise data owners, matched to the capability you need.
- Credential-verified experts
- Enterprise data partners
- Matched to your capability spec
corpuslab builds training and evaluation data for frontier AI labs, created by vetted experts or licensed from the enterprises that own it, with provenance on every record.
- Data
- Expert-made and licensed, rights-cleared at source.
- Research
- Evals and benchmarks that find where models break.
- Models
- Signal shaped for SFT, preference and RL.
- Infrastructure
- Provenance, privacy and delivery, built in.
01What we do
Human knowledge, made trainable. Expert data and licensed corpora for training and evaluating frontier models.
corpuslab connects the people and enterprises who hold scarce knowledge with the labs building frontier models, and turns that knowledge into data a model can learn from.
Expert data
Demonstrations, reasoning traces and preference judgments written by vetted specialists, not crowds approximating them.
Licensed corpora
Proprietary text, audio, video, code and workflow data from enterprises, licensed under clear rights and usage terms.
Evals & environments
Rubrics, verifiers, benchmarks and RL environments that measure real capability and reward the right behavior.
Provenance built in
Every record carries its source, consent and license, so you always know what your model learned from.
02Platform
From source to release. Four steps. One chain of custody.
02 License
License with clarity. Rights, consent and usage terms are set before a single record moves.
- Rights and consent on file
- Usage terms per model and run
- Revenue share for data owners
03 Verify
Verify every record. Expert review, automated checks and gold sets before anything ships.
- PII detection and redaction
- Multi-pass expert review
- Agreement and rubric scores tracked
04 Deliver
Deliver to your stack. Versioned datasets with data cards and provenance, shipped to your bucket or API.
- JSONL, Parquet or WebDataset
- Versioned releases and changelogs
- Provenance on every record
03Console
Every program, in one console. Experts, sources, reviews and releases, traceable end to end.
Traceable. Every record links to its source, consent and license.
Expert-led. Specialists write, review and score every task.
Private by default. Programs are isolated per customer.
Fast to first batch. Pilot data in days, not quarters.
04Who it's for
Built for every side of the data. Labs train on it. Enterprises license it. Experts create it.
Built to spec. Every program starts from your capability gap.
For AI labs
Signal you can train on. Scoped to the gaps your evals reveal.
Tell us where your model breaks. We scope the program, source experts and licensed data, and ship versioned datasets with provenance attached.
- SFT, preference and RL-ready formats
- Custom evals with sealed holdouts
- Custom data stays exclusive to you
Owner-first licensing. You approve every buyer and every use.
For enterprises
Your data, licensed on your terms. New revenue, without giving up control.
We value your data, de-identify it and license it to frontier labs under terms you approve. You keep ownership. We handle buyers, delivery and payouts.
- Rights, privacy and usage terms you set
- Revenue share on every license
- Approve each buyer; renew or end at term
Expert-first work. Clear rates, flexible hours, credit for quality.
For experts
Get paid for what you know. Remote work that shapes how frontier models reason.
Physicians, lawyers, engineers, scientists and linguists write and review the data frontier models learn from. Rates are set by domain and shown before you start.
- Rates posted upfront
- Choose your hours and projects
- Work matched to your field
Every record traced. Every contributor paid.
- Steps, source to release
- 4
- Record-to-license traceability
- 1:1
- Expert reviews per record
- 2×
- Scraped records
- 0
05Principles
Trust is the product. Six commitments behind every record we ship.
- 01
Consent before collection.
No record enters a corpus without documented consent and clear rights.
Applies to: Labs · Owners · Experts
- 02
Owners stay owners.
Licensing never transfers ownership. Buyers, terms and renewals stay your call.
Applies to: Owners
- 03
Experts are paid fairly.
Rates are posted upfront and reflect the expertise the work demands.
Applies to: Experts
- 04
Quality is measured, not claimed.
Every batch ships with its review scores, agreement metrics and known limits.
Applies to: Labs
- 05
Neutral by design.
We serve many labs and never share one customer's data or specs with another.
Applies to: Labs
- 06
Provenance is permanent.
Every record keeps its source, license and history for as long as it exists.
Applies to: Labs · Owners · Experts
Domains & modalities
From olympiad proofs to egocentric video.
- Mathematics
- Physics
- Chemistry
- Biology
- Medicine
- Radiology
- Pharmacology
- Law
- Finance
- Accounting
- Insurance
- Economics
- Software engineering
- Cybersecurity
- Mechanical engineering
- Education
- Reasoning traces
- Code
- Agents & tool use
- Computer use
- Audio
- Speech
- Music
- Video
- Images
- Documents & PDFs
- 3D & LiDAR
- Robotics
- Multilingual
- Creative writing
Engagements for labs and data owners.
Pilot
Prove fit on a scoped first batch.
Pricing: Fixed scope
- Capability-gap review with our researchers
- Rubric design and calibration
- Up to 500 expert tasks
- Quality report and data card
- Go/no-go in 3 weeks
Custom program
Dedicated experts and data for a capability target.
Pricing: Scoped per program
- Dedicated expert pods and reviewers
- SFT, preference and RL-ready data
- Custom evals with sealed holdouts
- Weekly versioned releases
- Program lead and research support
- Isolated workspace
Evals & environments
Measure capability before you train for it.
Pricing: Per benchmark or environment
- Expert-written tasks and rubrics
- Automated verifiers where possible
- RL environments for code, tools and computer use
- Re-runnable on every checkpoint
- Contamination-checked and private
Licensed corpora
Rights-cleared proprietary data from enterprise partners.
Pricing: Per license, term-based
- Text, audio, video, code and workflow data
- Consent and chain of title on file
- De-identified before delivery
- Usage terms per model or run
- Exclusive or non-exclusive
Data valuation
Learn what your data is worth to frontier labs.
Pricing: No-cost assessment
- Inventory of licensable sources
- Rights and consent review
- Demand check against open lab requests
- Indicative revenue range
- No data moves during assessment
Non-exclusive license
License to several labs. Keep full ownership.
Pricing: Revenue share
- You approve every buyer
- Usage terms per license
- De-identification handled
- Payout and usage reporting
- Renew or end at term
Exclusive license
One lab, premium terms, a defined window.
Pricing: Minimum guarantee + revenue share
- Single-buyer exclusivity for a fixed term
- Higher guaranteed minimums
- Tighter usage controls
- Usage audit rights
- Renewal option
Expert contributor
Get paid for what you know.
Pricing: Hourly or per task
- Rates shown before you start
- Remote, flexible hours
- Projects matched to your field
- Quality-based bonuses
- Payouts on a fixed schedule
Questions, answered
Everything labs, data owners and experts ask first.
Still have questions?Talk to usTwo places: work created by vetted experts under contributor agreements, and proprietary data licensed from the organizations that own it. Every record carries its source, consent basis and license terms. We don't sell scraped data of unclear origin.
We assess what's licensable, prepare it and offer it to labs under terms you approve. Each license pays a revenue share; exclusive deals add a guaranteed minimum. You see every buyer and every use, keep ownership, and decide on renewal at the end of each term.
Data is de-identified before delivery, using automated detection plus expert review for direct and indirect identifiers. Sensitive categories need the owner's explicit approval. Buyers receive only the approved, de-identified dataset, never raw source data.
A qualified expert writes each task, and a second expert reviews it. Each task is also checked against gold sets and automated validators. We track agreement, rubric pass rates and revision rates for each batch, and every delivery ships with a quality report.
Most programs begin with a scoped pilot. Sample data typically arrives within days and a first batch within weeks, depending on domain, volume and review depth.
JSONL, Parquet, CSV, WebDataset and Hugging Face-compatible datasets, or your own schema. We deliver to your S3, GCS or Azure bucket or through an API, with versioned releases, data cards and a provenance manifest.
Experts verify their identity and credentials, then pass domain assessments before joining a project. Their work is quality-scored on an ongoing basis. Pay is hourly or per task, set by domain, shown before you start, and paid on a fixed schedule.
Programs are isolated per customer. Access is least-privilege and logged, and data is encrypted in transit and at rest. Our controls are built for SOC 2 and GDPR-aligned workflows, and we sign DPAs and NDAs as standard.