-
v29: full model files sync — kernel, staging, eval, docs, training notebook
AVA
- kernel/ synced from canonical kaggle_upload source (matches Kaggle wyrm-500m-base)
- staging/ updated (synthase, PUP, SLA2)
- gaussian_head/ included in kernel
- training/wyrm_notebook_v29.py — real corpus loader, seq=1024, float32
- eval/wyrm_eval_victus.py — GPU eval with manual forward path
- docs: GLADIUS_TRAJECTORY, MODEL_CARD, SCALING_ANALYSIS, TRAINING_PLAN, ROADMAP
- kernel/NOTE.md — architecture decision log
- cleaned __pycache__ artifacts
- NO weights, code only
-
WYRM kernel v27 FINAL — live training version (443.7M base + Synthase + PUP + GaussianHead)
-
GLADIUS v5.0 — Cognitive kernel with Synthase depth attention, PUP uncertainty, Memory V2, multi-tokenizer architecture
