model: Qwen/Qwen3-4B
trust_remote_code: true
model_class: llm

template: qwen3_nothink

# Freeze Configuration
peft_config:
  name: freeze
  freeze_trainable_layers: 2      # Train the last 2 layers
  freeze_trainable_modules: all       # In these layers, train specific modules
  freeze_extra_modules: null      # Extra modules to train (e.g. embed_tokens, lm_head)

# Kernel Config
kernel_config:
  name: auto
  include_kernels: auto

# FSDP Config
dist_config:
  name: fsdp2
  dcp_path: null

### data
train_dataset: data/v1_sft_demo.yaml

### training
output_dir: ./outputs/test_freeze
micro_batch_size: 1
global_batch_size: 4
cutoff_len: 2048
learning_rate: 2.0e-5
bf16: false
max_steps: 10

### sample
sample_backend: hf
max_new_tokens: 128