RLHF and preference data, on your terms.
Alignment data only works when it reflects your rubric, not a vendor's generic guidelines. We produce preference datasets to your specification, in your domain, at the volume your training run needs.
What we produce
Preference data is judgment at scale. We work from your rubric and your examples so the judgments in the dataset are the ones you would have made.
- Comparison pairs. Chosen and rejected responses across your prompt distribution.
- Rankings. Full or partial rankings over multiple candidate responses.
- Critiques and rewrites. Annotated feedback showing why a response fails your rubric.
- Domain coverage. Technical, creative, safety, or product specific prompts sourced to match your use case.
Quality process
Work starts with a calibration batch against your rubric so you can check agreement before volume production. Ambiguous cases get flagged with notes rather than silently guessed.
Terms
Engagements are private, payment is in crypto only, and our price match guarantee means you will not find the same work cheaper elsewhere.
Common questions
How do you keep quality consistent at volume?
Calibration first, then production with spot checks. You review a sample before any large batch is produced.
Can you work from an existing rubric?
Yes, that is the ideal starting point. If you do not have one yet, we help you write it from a handful of examples.
What volumes can you handle?
From a few thousand pairs upward. Send your target volume to data@nexorpus.com and we will quote it.
Describe the data. We deliver it.
One email is all it takes. We reply with questions, samples, and a quote that beats the market.
data@nexorpus.com