model_name_or_path: /mnt/shared-storage-user/yuanfei/hf_hub/Qwen3-8B
trust_remote_code: true
stage: sft
do_train: true
finetuning_type: full
dataset: ft_sharegpt_top20pct_M1_LST_Qwen3_8B_rawfull_17langs_20260315_232435_en2xx_xx2en_merged_shuffled
template: qwen3
cutoff_len: 8192
preprocessing_num_workers: 16
overwrite_cache: true
packing: false
neat_packing: false
dataloader_num_workers: 4
output_dir: /mnt/shared-storage-user/yuanfei/yileitu/multidomain_rcot_translation/ft_outputs/Qwen3-8B_template-qwen3/ft_sharegpt_top20pct_M1_LST_Qwen3_8B_rawfull_17langs_20260315_232435_en2xx_xx2en_merged_shuffled_cutoff8192_20260421_131718
overwrite_output_dir: true
num_train_epochs: 1.0
per_device_train_batch_size: 4
gradient_accumulation_steps: 8
learning_rate: 1.0e-05
lr_scheduler_type: cosine_with_min_lr
warmup_ratio: 0.03
weight_decay: 0.01
max_grad_norm: 1.0
bf16: true
flash_attn: fa2
gradient_checkpointing: true
gradient_checkpointing_kwargs:
  use_reentrant: false
logging_steps: 10
save_steps: 50000
save_total_limit: 0
report_to: none
ddp_timeout: 180000000
eval_strategy: 'no'
lr_scheduler_kwargs:
  min_lr: 2.0e-06
plot_loss: true
dataset_dir: /mnt/shared-storage-user/yuanfei/yileitu/multidomain_rcot_translation/ft_scripts
deepspeed: /mnt/shared-storage-user/yuanfei/yileitu/multidomain_rcot_translation/ft_scripts/ds_z3_config.json