Human Data
Clinical Triage Reasoning - UK
Doctor-authored data for training models to decide what to ask, when to stop and when to escalate.
Why this data
The hard part comes before the answer.
Most medical datasets give a model the whole case and ask for an answer. Triage begins with missing information. The model must find the fact that changes care, recognise when a red flag makes further questioning unsafe, and stop once a safe route is known.
Clinical Triage Reasoning - UK turns doctor-authored symptom-checker pathways into auditable training records. The data is delivered with its clinical graph, explicit provenance and versioned transformations so teams can inspect the reasoning behind every training record.
Ask
Choose the missing fact that matters most.
Stop
Know when enough information is available.
Escalate
Act immediately when a critical feature is present.
Route
Select the appropriate UK care level without inventing a diagnosis.
Available configurations
One clinical source. Four training views.
Each configuration retains a link to the reviewed source graph and release manifest. Buyers can license the configuration required for evaluation, post-training or bounded production work.
Clinical graphs
Versioned questions, answer options, transitions, red flags and terminal care dispositions recovered from the source pathways.
- Designed for
- Graph policies, retrieval and auditable routing
Decision turns
Each visible conversation state paired with known facts, missing facts, acceptable next actions and prohibited actions.
- Designed for
- Supervised action tuning and process evaluation
Interactive trajectories
Complete and partial triage conversations grounded in a deterministic hidden patient state rather than an unconstrained simulator.
- Designed for
- Sequential training and reinforcement learning
Correction pairs
Reviewed safe actions contrasted with premature, irrelevant, delayed or unnecessarily escalatory alternatives.
- Designed for
- Preference and reward training
Production standard
Human provenance is the beginning, not the quality claim.
Human authorship alone is not a quality claim. Our release process keeps the source immutable, records every transformation and includes clinical review against the guidance version named in the manifest.
Recover
Restore an immutable source archive with record-level provenance.
Normalise
Rebuild pathway graphs and record every structural exception.
Review
Check clinical content against current UK guidance with qualified reviewers.
Transform
Generate linked graphs, turns, trajectories and correction data.
Version
Release signed manifests, schemas, data cards and change history.
Access
Licensed for the work you need to do.
The complete dataset is licensed rather than open sourced. Licences can cover evaluation, model training or bounded production use. Public materials include the schema, aggregate statistics, synthetic examples and dataset paper without exposing reconstructive records.
Need an edition for another health system? We can commission local care routing and independent clinical review rather than treating translation as localisation.
Request the dataset briefNo source records are published on this page. Dataset access does not establish that a trained model is safe or suitable for clinical deployment; that evidence must be established separately.