### model
model_name_or_path: /data/wyt/codes/DocDPO/sft/checkpoints_multilang/ted_base_balanced_en_zhdefr_320/sft
trust_remote_code: true

### method
stage: dpo
do_train: true
# finetuning_type: full
finetuning_type: lora
lora_rank: 8
lora_target: all
# deepspeed: /data/wyt/codes/DocDPO/sft/checkpoints_llama_factory/ds_z3_config.json  # choices: [ds_z0_config.json, ds_z2_config.json, ds_z3_config.json]

### dataset
dataset_dir: /data/wyt/codes/DocDPO/sft/data_multilang/red_multilang_base_balanced_en_zhdefr_320
dataset: dpo_en-zh_tool,dpo_en-zh_trans_base_sample,dpo_en-de_tool,dpo_en-de_trans_base_sample,dpo_en-fr_tool,dpo_en-fr_trans_base_sample
template: qwen
cutoff_len: 2560
# max_samples: 1000
overwrite_cache: true
preprocessing_num_workers: 16
dataloader_num_workers: 4

### output
output_dir: /data/wyt/codes/DocDPO/sft/checkpoints_multilang/ted_base_balanced_en_zhdefr_320/dpo/adapter
logging_steps: 5
save_steps: 200
plot_loss: true
overwrite_output_dir: true
save_only_model: false
report_to: tensorboard  # choices: [none, wandb, tensorboard, swanlab, mlflow]

### train
per_device_train_batch_size: 2
gradient_accumulation_steps: 4
# learning_rate: 1.0e-4
learning_rate: 5.0e-6
num_train_epochs: 2.0
lr_scheduler_type: cosine
warmup_ratio: 0.1
bf16: true
ddp_timeout: 180000000
resume_from_checkpoint: null

### eval
# eval_dataset: alpaca_en_demo
# val_size: 0.1
# per_device_eval_batch_size: 1
# eval_strategy: steps
# eval_steps: 500