# Layer-zero distillation: zero student layers 14-17 at init,
# constant LR 5e-7, 2000 steps. Aim: lower KL than the prior checkpoint
# despite the surgery.

[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    = 640
kl_start_pos   = 128
seed           = 42
shuffle_buffer = 10000

[train]
seed                 = 42
lr                   = 5.0e-7
schedule             = "constant"
warmup_steps         = 0
weight_decay         = 0.0
grad_clip            = 1.0
betas                = [0.9, 0.95]
eps                  = 1.0e-8
samples_per_step     = 8
micro_batch_size     = 8
max_steps            = 2000
grad_checkpointing   = true
attn_implementation  = "flash_attention_2"
student_dtype        = "bfloat16"
teacher_dtype        = "bfloat16"
mixed_precision      = "bf16"
kl_chunk_size        = 0
new_layer_lr_mul     = 1.0

[eval]
every_steps = 50
samples     = 64
seed        = 1234

[log]
wandb         = true
wandb_project = "distil-subnet97"
wandb_run     = "m-6a3lnzvb-zero14_17"
log_every     = 1
output_dir    = "./out/zero_14_17"

[init]
zero_layers        = [14, 15, 16, 17]
target_num_layers  = 32