# Cold start, 40 layers, low LR for original layers + 5x for the new ones.
# Lets the new layers wake up faster without disturbing the trained layers.

[model]
teacher    = "Qwen/Qwen3.5-35B-A3B"
student    = "Troiaaa/m-6a3lnzvb"
tokenizer  = "Qwen/Qwen3.5-35B-A3B"

[data]
dataset        = "karpathy/climbmix-400b-shuffle"
text_field     = "text"
min_chars      = 2560
max_seq_len    = 2048
kl_start_pos   = 128
seed           = 6767
shuffle_buffer = 10000

[train]
seed                 = 6767
lr                   = 1.0e-7
schedule             = "cosine"
warmup_steps         = 100
weight_decay         = 0.0
grad_clip            = 1.0
betas                = [0.9, 0.999]
eps                  = 1.0e-3
samples_per_step     = 4
micro_batch_size     = 4
max_steps            = 2000
grad_checkpointing   = true
attn_implementation  = "flash_attention_2"
student_dtype        = "bfloat16"
teacher_dtype        = "bfloat16"
mixed_precision      = "bf16"
kl_chunk_size        = 256
new_layer_lr_mul     = 5.0

[eval]
every_steps = 50
samples     = 500
seed        = 4242

[log]
wandb         = true
wandb_project = "distil-subnet97"
wandb_run     = "I_cold_paramgroups_grow40"
log_every     = 1
output_dir    = "./out/sweep/I_cold_paramgroups_grow40"

[init]
zero_layers        = []
target_num_layers  = 40