model: Qwen/Qwen3-0.6B
trust_remote_code: true
model_class: llm

kernel_config:
  name: auto

# FSDP Config
dist_config:
  name: fsdp2
  dcp_path: null

### data
train_dataset: data/v1_sft_demo.yaml

### training
output_dir: outputs/test_chunk_loss
micro_batch_size: 1
cutoff_len: 2048
# Maximum flattened token rows per logits/CE chunk; this is not the sequence length.
chunk_loss_size: 256
learning_rate: 1.0e-4
max_steps: 10

### sample
sample_backend: hf
max_new_tokens: 128
