A compact local Thai-and-English model pipeline that separates continual pretraining, supervised decision training, calibration, evaluation, and export into pinned configurations and artifacts.
Original by iApp TechnologyIntegrationsGitHub repositorySource reviewed
Before you dive in
What you’ll find in the original
Run continual pretraining, supervised decision tuning, and calibration as separate stages with separate outputs.
Version hyperparameters in YAML so a published checkpoint can be traced to one pipeline.
Evaluate the calibrated artifact before export and keep benchmark tables with the model card.
Worth knowing
The README estimates one H100 day for continual pretraining, half a day for supervised tuning, and hours for calibration. These resource estimates and the hosted endpoint were not verified by JevMade.