Linear-50M

Recipe: linear-50m-gdn2-delta-memory-read-mix-v5

GDN2 with a second delta-rule memory without global decay, read-aware gated mixing with per-channel agreement evidence, and an orthogonal memory fill-in term.

Training step: 80000

Validation loss: 2.9141

Downloads last month
286
Safetensors
Model size
49.6M params
Tensor type
BF16
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Datasets used to train User01110/50M-pro