qwen3_preprocess.yaml

### model
model_name_or_path: Qwen/Qwen3-4B-Instruct-2507
trust_remote_code: true

### method
stage: sft
do_train: true
finetuning_type: lora
lora_rank: 8
lora_target: all

### dataset
dataset: identity,alpaca_en_demo
template: qwen3_nothink
cutoff_len: 2048
max_samples: 1000
preprocessing_num_workers: 16
tokenized_path: saves/qwen3-4b/dataset/sft

### output (not used)
output_dir: saves/qwen3-4b/lora/sft
overwrite_output_dir: true