gladius-v2-kernel / MANIFEST.md
amuzetnoM's picture
WYRM v27 FINAL — locked training version (Day 58)
8035b6f verified

WYRM v27 FINAL — LOCKED VERSION

Date: April 8, 2026 (Day 58)

Status: TRAINING ON KAGGLE — DO NOT MODIFY

This is the exact version running on Kaggle T4 x2 right now.

Architecture

  • Total params: 565,816,475
  • Base kernel: 443,744,202
  • Synthase depth: 32,834,328 (6.89%)
  • PUP uncertainty: 6,406 (0.001%)
  • MultiEmbed: 33,560,576
  • Plug: 3,154,944
  • AuxHead: 32,770,048
  • GaussianHead: 19,745,971

Config

  • dim=1024, layers=24, heads=32, ff=4096
  • Batch: 4 x 4 accum = 16 effective
  • Sequence: 1024
  • Vocab: 16,000 (BPE)
  • Precision: float32
  • Optimizer: AdamW, 10 groups
  • Curriculum: language + math + cognition at multiple depths
  • SLA² on L0, Synthase depth profiles, PUP (disabled loss)

Files

  • wyrm_notebook_v27_FINAL.py — the Kaggle notebook
  • kernel-metadata.json — Kaggle push metadata
  • dataset/ — exact copy of kaggle_upload/ (kernel + staging + tokenizer)

Training Metrics (first 60 steps)

  • Loss: 11.31 → 4.81 (57% drop)
  • Task loss: 8.22 → 2.21
  • Best: 2.07
  • VRAM: 5.56 GB / 14.6 GB
  • Step time: ~27s
  • ETA: ~111h (15K steps)

Kaggle Details

  • Slug: amuzetnom02/wyrm-500m-interactive
  • Dataset: amuzetnom02/wyrm-500m-base
  • Accelerator: GPU T4 x2
  • Account: amuzetnom02