### model
model_name_or_path: Qwen2.5-7B
flash_attn: fa2


### method
stage: sft
do_train: true
finetuning_type: full
deepspeed: examples/deepspeed/ds_z3_config.json

### dataset
dataset: tower_zhen_cot_mt_gemini_distill.sft
template: chatml
cutoff_len: 12000
# cutoff_prompt_head: true
# max_samples: 16000
overwrite_cache: true
preprocessing_num_workers: 16
dataloader_num_workers: 4
# max_steps: 12540

# default_system: "You are Qwen, a virtual human developed by the Qwen Team, Alibaba Group, capable of perceiving auditory and visual inputs, as well as generating text and speech."

### output
output_dir: saves/qwen-7b/full/sft/v2
logging_steps: 1
save_steps: 2500
plot_loss: true
overwrite_output_dir: true
report_to: wandb  # choices: [none, wandb, tensorboard, swanlab, mlflow]

### train
per_device_train_batch_size: 4
gradient_accumulation_steps: 2
learning_rate: 1.0e-5
num_train_epochs: 1.0
lr_scheduler_type: cosine
warmup_ratio: 0.1
# warmup_steps: 300
bf16: true
ddp_timeout: 180000000
resume_from_checkpoint: null
# model_accepts_loss_kwargs: true

## eval
# eval_dataset: ted_yt_training_transcription.cot5.test.audio
# per_device_eval_batch_size: 1
# eval_strategy: steps
# eval_steps: 400