# 5M Monarch Mixer config — same param budget as 5m.toml
# 8 blocks (vs 6 for transformer) due to cheaper sequence mixing
# Monarch sequence mixer: 67K params/block vs 262K for attention

[model]
arch = "monarch"
embed_dim = 256
n_layers = 8
n_heads = 4           # unused by Monarch, kept for struct compat
head_dim = 64          # unused by Monarch
n_monarch_heads = 8
conv_kernel_size = 4
ffn_mult = 4
context_length = 256
dropout = 0.0
bias = false
weight_tying = true

[training]
optimizer = "adamw"
lr = 6e-4
min_lr = 6e-5
warmup_steps = 500
max_steps = 12305
batch_size = 32
grad_clip = 1.0
precision = "f16"
eval_interval = 500
eval_steps = 25
checkpoint_interval = 2000
seed = 42

[training.curriculum]
enabled = false

[training.coreset]
enabled = false

[data]
train_path = "../text-pipeline/output/train.txt"
val_path = "../text-pipeline/output/val.txt"
tokenizer_dir = "../text-pipeline/output"

[inference]
precision = "f16"
compile = false
temperature = 0.8
top_k = 40
max_new_tokens = 500