The Right Blend, Not the Biggest Pile
neuros-sourceweigher is a small, sharp service that decides how much each data source should
count when assembling a training mixture — the lever that keeps a foundation model from being
dominated by whatever dataset happens to be largest.
package map
neuros-sourceweigher — data-source weighting
A focused service that weights heterogeneous data sources for training mixtures, so a foundation model learns from the right blend rather than raw volume.
Comprises
Weigher
source-weighting logic
Service
API around the weigher
Role
Mixture control
balance datasets for pretraining
Quality over volume
down- and up-weight sources
What It Comprises
- Weigher — the source-weighting logic that scores and balances data sources.
- Service — an API wrapper so the weigher can be called from training and cloud pipelines.
How It Differs
It is the smallest package by surface area but a distinct concern: neuros-cloud moves and stores
data and neuros-neurofm trains on it, while neuros-sourceweigher governs the mixture —
turning "more data" into "better-balanced data" for pretraining.
Part of the neurOS-v1 platform.