Loading the leaderboard…
Loading the leaderboard…
03 · Dataset profile
Real-world Vietnamese medical speech recorded in actual doctor–patient conversations.
03.1 · Status
No numbers are published before the corpus is frozen — this page carries identity only, never preliminary stats.
03.2 · Onboarding
Six criteria every dataset clears before its first leaderboard — the same bar ViMedCSS cleared.
Redistribution rights documented up front — an open license, or a signed contract for held-out corpora like the private clinic set.
Where every clip came from, how transcripts were produced, and who verified them — published as part of the dataset profile.
Train, Valid, Test (and Hard where defined) frozen before the first run. Any revision creates a new dataset version — never a silent edit.
Systems that touched the labels — for example LLM-seeded transcripts — are declared and flagged permanently on that dataset's leaderboards.
References must score cleanly under the platform normalizer; orthographic quirks are documented and resolved before the first run.
WER, CER, CS-WER, N-WER, MTR, and RTFx must be computable from the references — or the gaps declared on the dataset profile.
03.3 · At launch
The same sequence every dataset follows — nothing is scored before the freeze.
Provenance, license, and split inventory go live on this page — frozen before any system is scored.
The platform harness runs the existing leaderboard systems against the new corpus, manifests and all.
Six metrics with bootstrap CIs and provenance chips from day one — comparable with every other dataset on the hub.