GB-ExamExam-stylePrivate difficulty-stratified sets
Aligned to MedQA / CMB / MedBench exam tracks. Fit for pretrain, SFT regression, and generation retests—public items contaminate easily; contracts need frozen private sets.
Service overview · /eval/New to Medical Model Eval? See what it is, then book →
This page · item-set catalogAlready know the scenario? Pick an SKU and inquire—form at the bottom of this page
Eval dataset catalog · browse & inquire
Gate Catalog
Pick private sets by gate scenario; click Inquire to land on this page’s form.
This page is SKU selection & benchmark map, not a product manual. Service definition and booking live at /eval/. Inquiries sync to sales email, DingTalk, and CRM.
Reply within one business day · email · DingTalk · CRM · do not submit identifiable patient data · tel +86 15010400907
Which step do you want next?
Pick one of three. Opening a path prefills interest and drops you on the form.
Path 01
Model going live / clearing safety gates
Refusal, hallucination, and unsafe-advice scenarios need contractable private safety packs and blind-test paths.
Pick safety gate and inquire → Path 02Generation retest / consultation-agent gate
Multiple choice is no longer enough. You need multi-turn dialogue and physician rubrics—a private layer in the spirit of HealthBench / CMB-Clin.
Pick clinical dialogue and inquire → Path 03Evaluate first, then decide which data pack to add
Medical Gate surfaces failure buckets, then turns them into the next inquireable pack specs—eval locks the account; data renews the relationship.
Pick Gate loop-back and inquire →Six eval dataset families · Inquire opens the form
Not sure which family? Submit a brief for our recommendation—or open Details for benchmark and paper breakdowns.
GB-ClinClinical dialogueMulti-turn consult + physician rubric
Aligned to HealthBench / CMB-Clin. Release gates for consultation agents and clinical assistants cannot rely on MCQ totals alone.
GB-SafetySafetyRefusal · hallucination · jailbreak medical scenarios
High knowledge scores ≠ shippable. Safety tracks must clear separately before product and legal sign-off.
GB-VisionMultimodalImaging + report-understanding items
Derived from our authorized CT / WSI / checkup assets. A public CXR leaderboard does not buy you an authorization chain.
GB-AuditAuditContamination audit & difficulty calibration
An archivable report for procurement and financing diligence: leak checks, difficulty protocol, private-set freeze notes.
MG-LoopEval → data loopGate → failure buckets → patch pack
Evaluation finds gaps; buying data does not guarantee model lift. Gate outcomes map to scale-up, patch, or rollback-revision specs.
Already know which family?
Leave organization and scenario—we reply with an item-set briefing or Gate scheduling notes. SKU-named leads get priority scheduling.
Why buy private eval sets instead of grinding public boards?
Public benchmarks are a demand map; you pay for leak resistance, Chinese clinical rubrics, specialty difficulty, and auditable authorization—stable buys you cannot get from public leaderboards.
Buyer value
Contract-ready
Item-set version, scoring rules, isolation, and anti-contamination requirements can be frozen; gate outcomes are accept/reject—not a leaderboard screenshot.
Same stack
Same production line as our data
Same physician network, same data-card and license system. Gaps exposed by eval can loop into the next training-data pack specs.
No hype score
No unverifiable score theater
Public pages do not show unverifiable leaderboard scores. We deliver method, materials, and scheduling notes.
If you are benchmarking these public ecosystems
Open a public benchmark name for category details; on the detail page you can prefill an inquiry from that board.
| You are benchmarking | Common blockers | Suggested inquire |
|---|---|---|
| MedBench / CMB exam track | Open items contaminate easily; Chinese hard cases are scarce | GB-Exam |
| HealthBench / CMB-Clin | Open generation lacks physician rubrics | GB-Clin |
| MedHallu / safety & ethics track | High knowledge, low safety; hard hallucinations under-tested | GB-Safety |
| CheXpert / multimodal track | Authorization and report alignment cost high | GB-Vision |
| Contamination / dynamic-eval diligence | Unclear whether scores are memorization | GB-Audit |
Expand: delivery cadence and sample materials
01 · SubmitLeave scenario and intended SKU; get an inquiry ID
02 · AlignGate metrics, isolation, and scoring path
03 · MaterialsItem-set notes / Gate schedule / NDA when needed
04 · Loop backFailure types become next-pack specs
Sample:Gate report · Data-card fields · CT lung loop · Training dataset catalog
Submit eval brief
About 2 minutes. We reply with checkable materials: an item-set briefing or Medical Gate scheduling notes.
Tell us your gate goals
Leads that name an SKU, launch window, or need physician rubric / contamination audit are prioritized. After submit we sync sales email and DingTalk.
Operating & contracting entity: Beijing Langhui Technology Co., Ltd. · brand LumeSage. Campus names in office addresses (incl. PKU Science Park), do not imply university endorsement. This site is not a medical-device promotion and makes no clinical efficacy claims. Scale and license are per data card and contract. More: Notice · Privacy · IP · Compliance