SFT
Expert demonstrations and step-by-step trajectories for supervised fine-tuning.
Expert tasks & RL data
Tasks and trajectories from working professionals.
Realistic, multi-step tasks built and solved by practitioners — with demonstrations, reward signals, and critique that teach agents to do real work.
What it is
Real work, step by step.
Agents learn real work from people who do it. Vetted professionals design realistic tasks, solve them, and judge the results.
- Realistic task environments and prompts
- Expert demonstrations and step-by-step trajectories
- Reward rubrics, preference pairs, and critiques
- SFT and RLHF/RLAIF-ready formats
- Coverage across professional domains
Real tasks, real practitionersExpert tasks
Who it’s for
Built for teams training agents that do real work.
RLHF / RLAIF
Preference pairs, reward rubrics, and scores for reward modeling and policy optimization.
Agent environments
Realistic multi-step tasks and prompts across professional domains.
How it works
Designed, solved, and judged by experts.
Scope
We agree domains, task types, signals, and format with you on a short scoping call.
Build & solve
Vetted practitioners author realistic tasks and record demonstrations and trajectories.
Judge & QA
Experts write rubrics, preference pairs, and critiques, then multi-pass QA reviews every record.
What’s in a record
Every task ships with the signal to learn from.
- Task prompt and environment context
- Expert demonstration and trajectory steps
- Reward rubric and scores
- Preference pairs and written critiques
- Domain and expert-role metadata
- Review passes and approval status
FormatsJSONL, parquet, custom
SignalsPreferences, rubrics, scores
DomainsFinance, health, legal, code, STEM
ExpertsVetted for credentials
Every delivered recordYour schema
- Source
- Domain, task, modality, expert role & experience
- Content
- Media or text, with timestamps where relevant
- Labels
- Steps, actions, objects, intents, rubrics, scores
- Expert notes
- Rationale, technique, edge cases
- Privacy
- Face blur & PII redaction flags
- Consent
- Contributor agreement & site/owner consent
- QA
- Review passes and approval status
- Format
- JSON · JSONL · CSV · parquet · COCO
FAQ
Expert RL & agent data: common questions
What is expert RL data for agents?
Realistic, multi-step tasks built and solved by working professionals, with demonstrations, step-by-step trajectories, reward signals, and critique that teach agents to do real work.
What signals do you provide?
Reward rubrics, preference pairs, scores, and written critiques from vetted experts.
Which formats are supported?
SFT and RLHF/RLAIF-ready formats in JSONL, parquet, or a custom schema, delivered to S3, a private Hugging Face repo, or your bucket.
Which domains do you cover?
Finance, healthcare, legal, software engineering, mathematics, science, and operations.
Who are the experts?
Practitioners vetted for their credentials in the relevant domain, not generic crowd workers.
How do we get started?
Start with a short scoping call, then a pilot sample set in your format, then scale with agreed throughput and QA targets.
Start with a sample
See the data before you commit.
Tell us what your model needs to learn. We’ll scope a pilot and send a sample set in your schema.