RLHF and alignment data for language models
Preference, safety and evaluation data, ranked and checked by people who know the subject of the question. For Frontier model labs and open-source fine-tuners.
The problem
Preference data is only as good as the people ranking it. Rankers who do not know the subject reward answers that sound right, which trains a model to sound right rather than be right.
Work we take on
| Task | What you receive |
|---|---|
| Response ranking | Ranked model answers with a short reason for each preference. |
| Adversarial red-teaming | Prompts that break the model, grouped by failure type, with the failing responses. |
| Safety calibration | Labels for refusals that should not happen and answers that should have been refused. |
| Hallucination reduction | Claims in model answers marked as supported, unsupported or wrong, with the evidence. |
How a batch moves
- Pre-label
- Domain-tuned LLMs draft first-pass rankings and flag likely unsupported claims for review. Live
- Who reviews it
- Subject experts for the fields your model must answer in, plus generalist reviewers for tone and safety. We are recruiting them now. Next
- Agreement check
- Blind multi-review on ambiguous items, tracked with Fleiss’ kappa and Cohen’s kappa. Next
- Status
- Taking pilot requests. Pre-labelling and security controls run today. Expert review starts with the first pilots.
Engagement models that fit
- Managed annotationPer unit, with service levels tied to verified accuracy thresholds. Planned.
- Dedicated expert podsA monthly contract for a team of industry analysts and engineers who work only on your data. Planned.
- Evaluation setsAn annual licence for several seats. Planned, sets not yet available.
Request a pilot for RLHF and alignment
Pilots start once our first expert reviewers are in place. Send a request now and we will reply by email.