SHOREKEEPER / configs /model.yaml
geoore's picture
Restructure to src/ layout with attention, per-layer MoE, and working chat
73400c8
model:
name: "SHOREKEEPER-4B"
version: "1.0.0"
dim: 3072
n_layers: 28
n_heads: 24
n_kv_heads: 6
head_dim: 128
vocab_size: 50304
seq_len: 8192
n_experts: 12
n_activated: 2
expert_dim: 2048
experts:
router: "Sentinel"
members:
- name: "Asmoday"
role: "code"
specialization: "python_development"
- name: "Istaroth"
role: "systems"
specialization: "os_networking"
- name: "Ronova"
role: "reasoning"
specialization: "math_logic"
- name: "Naberius"
role: "memory"
specialization: "retrieval"
- name: "Phanes"
role: "creation"
specialization: "writing"
- name: "Barbeloth"
role: "analysis"
specialization: "data_patterns"
- name: "Tacet"
role: "silence"
specialization: "filtering"
- name: "Abby"
role: "empathy"
specialization: "user_context"
- name: "Reindoter"
role: "validation"
specialization: "testing"
- name: "Zestial"
role: "vision"
specialization: "visualization"
- name: "Alice"
role: "exploration"
specialization: "novelty"
- name: "Rover"
role: "execution"
specialization: "terminal"
rope_theta: 1000000.0
quantization:
bits: 4
type: "nf4"
double_quant: true
compute_dtype: "bfloat16"