Linked at the source. Structured for training and evaluation.
Every record has receipts.
Consultation audio, transcripts, notes and documents, with their source connections intact.
Someone was in the room.
Clinical notes reviewed by physicians within the care workflow.
Travel in time.
Encounters, documents and structured FHIR records connected over time.
Speak FHIR.
One ontology for clinical records, care delivery and insurance operations.
Bring your own question.
Define the population, modalities and review criteria. We shape collection and recurring delivery around your needs.
DELIVERY
FHIR R5 + Croissant. ICD-10 and UCUM where present. Modalities and follow-up vary by cohort.
PREPARATION
Identifiers removed and checked before delivery. Speaker labels, codes and extracted values are model-produced; not all are independently validated. Samples are for evaluation only.
Your next dataset has a day job.
Source datasets for your AI lab customers, with agreed cohorts, quality criteria and delivery cadence.
Charcot Gym
727 public tasks · 5 families · 315 Hard · 126 image · 189 private
01
Document
HbA1c · 9.8%
02
Agent actions
fhir.create(Observation)
03
Final record
LOINC 4548-4 · UCUM %
04
Deterministic verifier
code · value · unit · date
Scored on the final record state. Offline. No LLM judge.
Charcot Gym turns our data engineering into synthetic tasks, RL environments and verifiable outcomes.
14 models on 99 private tasks. Three attempts per task. November 2026.
Overall is the mean over scored tasks. Models with a footnote did not complete every cell and are not directly comparable. The top three overlap within their confidence intervals.
¹ Image input failed through the provider API, so Hard image was not run.
² Most clinical coding attempts were refused by the provider’s safety filter and scored 0.
³ Hard tasks timed out or failed on the provider API.
⁴ Text-only model; Hard image not supported.
Llama 4 Maverick was withdrawn because it did not emit structured tool calls.
Our models
01 / CRS-4B
A 4B model went to the Gym and came back with a thing for healthcare admin. Odd hobby. Useful one.
CRS-4B · Qwen3.5-4B base → fine-tuned
99 private tasks · 3 attempts per task. Frontier models: 0.76–0.89.
02 / MARIMBA
System One for speech. Still backstage. You’ll hear more when it’s ready.
Small models. Unreasonably specific interests.
CRS-4B is fine-tuned only on public synthetic Gym episodes and evaluated on the private split.
Coming in v2
01
Source difficulty levels
CSV, HL7 v2, broken-table PDFs, superseded reports, non-standard units, ambiguous dates and same-name patients.
02
Difficulty bands
Difficulty bands for every task.
03
Clinical scores
Calculators implemented from the primary publications, including the correct refusal when inputs are missing.
04
Web EHR computer use
A synthetic web EHR with screenshots, coordinates, an accessibility tree and human approval steps, scored on the final FHIR state.
05
A larger public task set
Published as a versioned dataset.
Tell us what data you need. We’ll shape the cohort, format and delivery around it.