// MODEL
Occam 1.5B is a 28-layer transformer trained in six languages on the SMET architecture. The design trades parameter count for processing depth -- instead of scaling memory to store more facts, the model uses its layers to evaluate what it knows and regulate its own output accordingly. An internal certainty signal measured across all layers tells the model whether it has the knowledge to answer a given query before it generates. When it doesn't, the model can refuse or trigger retrieval autonomously.
// LIVE TRAINING DATA
● Occam 1.5B Training Monitor
Real-time epistemic signals, loss trajectory, per-layer GSI, SMET diagnostics
// TRAINING REPORTS
Periodic reports published at training milestones. Each checkpoint includes a training diagnostics report and a 100-question benchmark across six languages. HellaSwag evaluation added from step 27,000 (1× Chinchilla).