model: "models/Qwen3.6-35B-A3B"
data: "data/micro-kiki/cpp"
train: true
fine_tune_type: lora
iters: 500
batch_size: 1
learning_rate: 1e-5
adapter_path: "output/micro-kiki/lora-qwen36-35b-v4-sota/cpp"
max_seq_length: 512
num_layers: 32
steps_per_report: 50
steps_per_eval: 100
grad_checkpoint: true
lora_parameters:
  rank: 16
  alpha: 16.0
  dropout: 0.0
  scale: 20.0