Work with us

Teach the model what better means.Judged by experts in the field.

Saolabs produces alignment data from expert judgement: rankings, preference data, reference answers and reasoning. It gives models a clear signal about which answers are safer and more correct.

What alignment data is

Alignment data teaches a model which behaviour you want. It usually takes the form of human judgements about model outputs, plus examples of what a good answer looks like.

Saolabs produces four main kinds, each written or judged by a vetted specialist in the relevant field.

  • Rankings of several model responses to the same prompt
  • Pairwise preference data for RLHF-style and related training methods
  • Expert reference answers that show the target behaviour
  • Step-by-step reasoning that explains how an expert reaches a judgement

Why expert judgement matters here

Preference training pushes a model toward whatever the judges prefer. If judges reward the confident, fluent answer over the careful, correct one, the model learns that too.

Experts prefer the answer that is right and safe to act on. In medicine, that might be the response that says not to double a missed dose and to check with a doctor, rather than the one that simply says yes.

Rubrics and rationales

We agree the criteria for each judgement with your team before production: correctness, safety, helpfulness, appropriate uncertainty, or the trade-offs specific to your model.

Experts can attach a short rationale to each judgement. That makes the data auditable, helps you find inconsistent preferences, and gives you material for training reward models or writing future guidelines.

Handling disagreement

Qualified experts do not always agree, and that is useful information. Low agreement on an item usually means the prompt is ambiguous, the rubric is unclear, or the question has more than one defensible answer.

Rather than hide that, we aim to surface it. Items where experts split can be reviewed, reworded or delivered with a note, so your team decides how to treat them in training.

How an alignment-data project runs

We scope the domains, behaviours and data format with you, and draft rubrics together. A sample set comes first, so your team can check the judgements against its own view of better before production.

We then match vetted experts and produce the data, with expert review and quality checks. Delivery comes in your agreed format with notes on known limits, such as low-agreement items or areas with thinner coverage.

Questions, answered.

What is AI alignment data?

AI alignment data is human feedback used to steer a model toward desired behaviour. It includes rankings and preference judgements on model outputs, reference answers and written reasoning.

What is the difference between preference data and SFT data?

SFT data pairs prompts with reference answers that a model learns to imitate. Preference data records which of several responses a judge prefers, and is used to train the model or a reward model to favour better answers. Many training pipelines use both.

Why use domain experts for RLHF preference data?

Preference data trains a model toward whatever the judges reward. Domain experts reward answers that are correct and safe to act on, rather than answers that only sound confident, which matters most in fields like medicine, law and finance.

How do you handle disagreement between annotators in preference data?

Disagreement between qualified annotators usually signals an ambiguous prompt, an unclear rubric or a question with several defensible answers. Good practice is to measure agreement, review the split items and either fix them or deliver them with a clear note.

What formats does alignment data come in?

Alignment data is typically delivered as ranked lists or pairwise comparisons of model responses, often with rationales, plus reference answers and reasoning traces. Saolabs agrees the exact schema with each client so it fits their training pipeline.

Every safe modelhas an expert behind it.

Work with usJoin as an expert