Expert SFT Data: Using Doctors, Lawyers and Engineers to Write Training Data

Why Generalist SFT Data Fails in Specialist Domains

Supervised fine-tuning works by showing a language model examples of the responses it should learn to produce. The model generalises from those demonstrations. The problem with using general-purpose annotators for domain-specific tasks is that what they produce — however grammatically polished — can encode the wrong reasoning, incorrect terminology, or advice that no practitioner would give.

A 2023 study published in JAMA Internal Medicine found that general-purpose LLMs produced factually incorrect responses to clinical questions in 26% of cases when responses were reviewed by practising physicians. More significant was the finding that the errors were not random: they clustered around subtleties that a non-expert annotator would not recognise as errors at all. The model had learned to produce confident-sounding but incorrect clinical reasoning — exactly the output its SFT data had demonstrated.

This is the core problem with generalist SFT for high-stakes domains. It does not produce obvious nonsense. It produces responses that are wrong in ways that require domain expertise to identify. Fixing this after deployment is expensive and reputation-damaging. The correct intervention is upstream: use domain experts to write the demonstrations in the first place.

What Counts as an "Expert" for SFT Purposes

The definition of "expert" is task-specific, not credential-specific. A practising GP is an expert for primary care clinical reasoning tasks. A medical student in their third year is not — their knowledge may be roughly similar on some topics but their clinical judgement is not calibrated. A tax partner at a Big Four firm is an expert for corporate tax SFT. A paralegal with three years of experience in family law is an expert for residential property conveyancing. An expert is someone whose domain knowledge is current, practically tested, and sufficient to recognise errors when they see them.

For expert SFT and RLHF data projects, the common expert categories AI Taggers works with include:

• Medical: GPs, specialists, nurses, pharmacists, dentists, allied health practitioners — matched to the specific clinical scope of the model.
• Legal: Solicitors and barristers by practice area — commercial, family, criminal, tax, intellectual property. Australian jurisdiction expertise where local law applies.
• Engineering: Civil, structural, mechanical, electrical, and software engineers at practitioner level (not graduate), matched to the engineering domain of the product.
• Financial: CFAs, CPAs, financial planners (RG146-compliant in Australia), commercial bankers, credit analysts.
• Scientific: PhD-level researchers or senior industry scientists for bench chemistry, materials science, and environmental science tasks.

Credential verification is a prerequisite, not an optional step. For medical and legal experts, AI Taggers checks registration with AHPRA (Australia) or the relevant national body before any expert accesses a production annotation task.

A Representative Expert SFT Deployment: Healthcare AI

To make the quality case concrete, consider the typical pattern from enterprise healthcare AI deployments. A team building a clinical decision support assistant typically starts with a base LLM fine-tuned on general medical text. They run it against a physician-reviewed evaluation set and find an error rate of 20–35% on clinical reasoning tasks — the exact range documented in published clinical LLM benchmark studies including work from Stanford AIMI and NEJM AI.

After an expert SFT intervention — typically 2,000–5,000 GP-written demonstrations covering the highest-volume clinical query types — error rates on the same evaluation set drop to 8–12% in most published cases. The residual errors concentrate in edge-case clinical scenarios outside the SFT coverage, which informs the next round of data collection.

Critically, the cost-per-correct-answer calculation shifts in favour of expert data at this error rate differential. If an incorrect response costs ten times as much to address as a correct one (clinician review, override workflows, retraining), then halving the error rate more than pays for the premium on expert annotation even before counting the reputational costs of clinical AI errors.

How Expert SFT Projects Are Quality-Controlled

Quality control for expert SFT is structurally different from standard annotation QA. You cannot use a cheaper annotator to review an expert's work — the reviewer needs at minimum equivalent domain expertise. This requirement doubles or triples the senior headcount relative to a generalist annotation project.

The standard protocol at AI Taggers for expert SFT involves:

1. Expert onboarding and calibration: A task-specific calibration session where experts complete 20–30 practice items against a gold-standard answer set prepared by a senior domain reviewer. Experts who fall below 80% agreement on the calibration set are not progressed to production.
2. Dual-expert review for high-risk domains: For medical and legal tasks, every fifth production item is independently written by a second expert and the two responses are compared. Systematic divergence triggers a rubric review, not rejection of the individual expert.
3. Rolling inter-annotator agreement (IAA) monitoring: IAA is tracked on a weekly basis using Cohen's kappa. Expert cohorts below a kappa of 0.75 on rating tasks (or below 0.70 on generation tasks) trigger recalibration before the data reaches final QA.
4. Senior QA audit: A senior domain reviewer — typically a specialist one level above the production expert cohort — samples 5–10% of completed items and reviews them against the rubric. QA findings are returned to experts as coaching, not rejection counts.

This structure means expert SFT data typically arrives with documentation: IAA scores, calibration pass rates, QA audit findings, and provenance logs. For medical AI destined for FDA submission, that documentation is not optional — it is a regulatory requirement under 21 CFR Part 11.

Cost Drivers and What to Budget

Expert SFT costs more for four compounding reasons: expert hourly rates, lower throughput (experts write more slowly and carefully than generalists), senior QA overhead, and higher task design cost (rubrics for expert tasks must be clinically or legally precise, which requires expert input to draft).

Typical per-response ranges for production expert SFT in 2026:

These ranges are for well-scoped, production-quality work with proper QA. For minimum viable expert SFT — 1,000 responses, single domain, pre-vetted expert panel — budget AUD $50,000–$100,000 depending on domain and response complexity. Pilot projects of 100–200 responses to validate rubric and panel quality are available from AUD $8,000–$15,000.

How to Scope an Expert SFT Project

Scoping starts with the failure mode, not the dataset size. Identify the specific query types where your current model produces clinically or legally problematic outputs. A systematic evaluation using your own domain experts against a sample of 200–500 model responses will surface the failure clusters. Those clusters become the annotation brief.

Once the failure modes are mapped, estimate coverage volume. For each identified failure cluster, you need enough expert demonstrations to shift the model's behaviour — typically 200–500 per cluster for a well-scoped, focused fine-tuning run. Attempting to cover every possible domain query in one SFT round is rarely cost-effective: it spreads the expert budget too thin to achieve measurable change in any single failure mode.

The right sequence is:

1. Run a structured evaluation to identify failure clusters.
2. Prioritise clusters by risk and query frequency.
3. Commission a scoped pilot of 100–200 expert demonstrations against the top cluster.
4. Fine-tune and measure against your evaluation set.
5. Iterate across remaining clusters in order of priority.

This iterative approach also surfaces rubric problems early, before they propagate across a large dataset. It is more time-efficient than a single large data collection round that turns up poor quality only at final QA.

Starting with a Free Pilot

Before committing to a full expert SFT dataset, it is worth running a small pilot — 50–100 responses — to validate three things: that the rubric is precise enough to guide expert writers consistently, that the expert panel produces responses at the required quality level, and that the fine-tuned model shows movement on your evaluation set.

AI Taggers offers a free sample batch for new expert SFT and RLHF data projects so teams can verify data quality and rubric fit before any commitment. Contact us with your domain, target query types, and any existing evaluation data, and we will scope a pilot proposal within one business day.

Frequently Asked Questions