TestSubject-2 / README.md
CyanMonkey's picture
step 92000 | ppl=10.62
8284cad verified
|
Raw History Blame Contribute Delete
427 Bytes
metadata
language: en
tags:
  - causal-lm
  - gqa
  - qk-norm
  - rope
  - swiglu
  - refresh-gate
  - rose-x1
license: apache-2.0

TestSubject-2

Rose X1.5 SLM (98.8M params).

Architecture

Property Value
Layers 22
Hidden 576
Heads 9 (kv=3)
QK Norm True
XSA True
Refresh Gate [6, 13]
Params 98.766M

Training

  • Tokens: 48,210,944,000
  • Val PPL: 10.62
  • Optimizer: muon_adamw