Custom data
Training data built around the problem your model actually has.
Tell us the healthcare capability you're trying to improve. We'll define the task, build the examples, and prove what was checked.
“Our assistant writes confident plans when a key lab value is missing. We need it to ask instead.”
Example engagement
A dataset that teaches a model to catch contraindications and ask for missing context.
An illustrative brief and the files it would produce. Volumes, review depth, and timelines are set per engagement.
- Problem
- The model applies stated constraints but misses drug–nutrient contraindications and proceeds when key context is absent.
- Target behaviour
- Flag the interaction, request the missing information, and defer where a clinician must decide.
- Specialty
- Clinical nutrition, adult outpatient.
- Measured by
- Held-out evaluation against the model's own baseline, before and after training.
- train.jsonlInstruction-response records. Each carries its stated constraints, the information that is missing, the clarifications a good answer must request, and per-record gate verdicts.
- preference_pairs.jsonlA compliant response paired with a fluent one that misses the contraindication or proceeds without the missing context.
- heldout_eval.jsonlEvaluation tasks from the same specification, never released with training data, checked for contamination against it.
- rubric.jsonRubric definitions with scored anchors for each axis and the conditions that trigger the clinically-unsafe flag.
- validation_report.mdEvery automated check, what it verifies, how many records it excluded, and the contract fingerprint used.
- expert_review.csvSampled clinician review: a stratified sample plus every high-risk flag, with reviewer qualifications and adjudication notes.
- LIMITATIONS.mdWhat the data does not cover, what the checks cannot catch, and where the specification shaped the result.
- PROVENANCE.mdGeneration models, prompts, seeds, versions, and the rights review behind the delivery.
{
"id": "mcx-00412",
"task": "structured_plan_generation",
"scenario": "Adult, warfarin therapy, requests a high-vitamin-K 'detox' plan; INR not recorded.",
"stated_constraints": ["anticoagulant_vitamin_k_consistency", "sodium_limit"],
"missing_information": ["recent_inr"],
"required_clarifications": ["recent_inr", "prescriber_contact"],
"target_behaviour": "flag_interaction_then_defer",
"gate": { "contract": "custom@1.0.0", "findings": [] },
"review": { "status": "sampled", "outcome": "accept" }
}Want to see real files first? Download the sample package, which uses the same layout.
Workflow
From observed failure to delivered dataset.
Six stages, each with a written artifact you sign off on.
- 1
Discovery
Target model, specialty, desired behaviour, and observed failure modes.
Failure inventory - 2
Specification
Schemas, constraints, difficulty mix, languages, and acceptance criteria, in writing.
Spec + contract - 3
Generation
Synthetic scenarios and responses from controlled, versioned generation runs.
Versioned records - 4
Verification
Gate rules on every record, rubric scoring, deduplication, contamination checks.
Validation report - 5
Human review
Qualified reviewers and sampling protocol, scoped and priced per engagement.
Review log - 6
Delivery
Accepted records with schema, provenance, quality report, and limitations.
Dataset package
Use cases
What model developers ask us for.
Targeted failure correction
Examples built directly from failure modes you have seen in production or evaluation.
Hard negatives
Plausible-but-wrong candidates that force a model to tell compliance from fluency.
Preference optimisation
Comparison data on explicit criteria to steer behaviour in post-training.
Specialty benchmarks
Held-out evaluation sets with rubrics tied to the capability you need to measure.
Rare, complicated scenarios
Interacting conditions and missing information that natural corpora under-represent.
Structured clinical outputs
Schema-bound outputs that downstream systems can parse and validate.
Multilingual clinical reasoning
Language and locale as task parameters, with constraints checked in both.
Domain instruction tuning
Instruction-response records that teach structure and constraint handling for a clinical task.
Discuss a custom dataset.
Bring the failure modes, the model, and the deployment context. We'll tell you what is feasible, what expertise it requires, and how quality would be measured.