train_full_deepspeed.yaml

model: Qwen/Qwen3-0.6B

model_class: llm

template: qwen3_nothink

kernel_config:
    name: auto
    include_kernels: auto

dist_config:
    name: deepspeed
    config_file: examples/deepspeed/ds_z3_config.json

### data
train_dataset: data/v1_sft_demo.yaml

### training
output_dir: outputs/Qwen3-0.6B-deepspeed
micro_batch_size: 1
cutoff_len: 2048
learning_rate: 1.0e-4
bf16: true
max_steps: 10