JevMade field notes / Inference and validation guide
KaLM-Jev
KaLM-Jev explains local Choice, Score, and Noul aggregation, cache behavior, precision-specific validation receipts, and why its default probabilities are not correctness estimates.
Original by KaLM-EmbeddingIntegrationsGitHub repositorySource reviewed
Before you dive in
What you’ll find in the original
Preserve full-precision outputs and execution metadata even when README examples round probabilities.
Document how softmax, expected-level scores, and true/false margins produce different typed outputs.
Treat document-cache sizing, precision, and request limits as part of the reproducible runtime.
Worth knowing
The displayed BF16 examples are explicitly illustrative and use default uncalibrated aggregation; they do not establish overall accuracy. Primary validation used an H100 MIG instance.