badam_mode: layer
badam_switch_interval: 100
badam_switch_mode: ascending
badam_update_ratio: 0.1
bf16: true
cutoff_len: 4096
dataset: smoltalk_chinese
dataset_dir: data
ddp_timeout: 180000000
deepspeed: cache/ds_z3_config.json
do_train: true
eval_steps: 5000
eval_strategy: steps
finetuning_type: full
flash_attn: fa2
gradient_accumulation_steps: 4
learning_rate: 0.0003
logging_steps: 5
lr_scheduler_type: cosine
max_grad_norm: 2.0
max_samples: 10000000
model_name_or_path: deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B
neat_packing: true
neftune_noise_alpha: 5
num_train_epochs: 2.0
optim: adamw_torch
output_dir: saves/DeepSeek-R1-Distill-Qwen-1.5B/full/train_2025-01-23-00-42-56
packing: true
per_device_eval_batch_size: 4
per_device_train_batch_size: 4
plot_loss: true
preprocessing_num_workers: 16
report_to: none
save_steps: 5000
stage: sft
template: deepseekr1
trust_remote_code: true
use_badam: true
val_size: 0.001
warmup_steps: 100