One protocol per task
Ranks compare the same data and settings. Speech, translation, and language models have separate tracks. Filters do not rerank models.
Loading models
Reading the published snapshot.
KALENJIN MODELS
Speech recognition
Explore models that transcribe Kalenjin speech.
Chamgei edited speechLoading evaluation
Reading the published protocol and result snapshot.
Only common-protocol evaluations determine rank.
About the Chamgei evaluationDataset & protocol
Results are specific to the published protocol and submitted, edited references. A computational audit checks arithmetic and recorded artifacts; it is not peer review or independent native-speaker adjudication.
The five evaluation recordings, edited references, and all seven models’ responses are available on the Listen & compare page. Training overlap and speaker independence remain unknown; this is not a demonstrated unseen or blind test. All inference and scoring run on the owner’s DGX Spark.
Read the protocol and access limits ↗How to read this leaderboard
Methodology ↗Read the audit scope
Computational checks are not peer review. A dash means no eligible score; unavailable and excluded models retain a visible reason.
Author claims are context
Reported scores live in expandable notes. They never affect rank. Maintainer-owned models follow the same rules.