### model
model_name_or_path: /shared_workspace_mfs/ximing/Qwen3-8B-Base
trust_remote_code: true

### method
stage: sft
do_train: true
finetuning_type: full
deepspeed: /shared_workspace_mfs/ximing/LLaMA-Factory/examples/deepspeed/ds_z3_offload_config.json

### dataset
dataset: test_10_samples
template: qwen3_nothink
cutoff_len: 8192  # 降低 cutoff_len，减少内存使用
rope_scaling: yarn
overwrite_cache: true

preprocessing_num_workers: 4  # 减少 worker 数量
dataloader_num_workers: 2

### output
output_dir: /shared_workspace_mfs/ximing/sft_test_10_samples_debug
logging_steps: 1
save_strategy: steps
save_steps: 10
plot_loss: true
overwrite_output_dir: true
report_to: none  # 不使用 wandb，减少开销

### train
per_device_train_batch_size: 1
gradient_accumulation_steps: 4  # 减少梯度累积步数
learning_rate: 5e-5
num_train_epochs: 1  # 只训练 1 个 epoch 用于测试
lr_scheduler_type: cosine
warmup_ratio: 0.1
bf16: true
ddp_timeout: 180000000
enable_liger_kernel: false  # 关闭 liger kernel，避免潜在问题
use_unsloth_gc: false  # 关闭 unsloth gc，避免潜在问题
flash_attn: fa2