bf16: true
cutoff_len: 4096
dataset: train_hal_detection
dataset_dir: data
ddp_timeout: 180000000
deepspeed: cache/ds_z2_config.json
do_train: true
finetuning_type: full
flash_attn: auto
gradient_accumulation_steps: 8
learning_rate: 5.0e-06
logging_steps: 1
lr_scheduler_type: cosine
max_grad_norm: 1.0
max_samples: 100000
model_name_or_path: meta-llama/Meta-Llama-3-8B
num_train_epochs: 3.0
optim: adamw_torch
output_dir: saves/Llama-3-8B/full/train_2024-11-22-23-46-34
packing: false
per_device_train_batch_size: 1
plot_loss: true
preprocessing_num_workers: 16
report_to: none
save_steps: 1000
stage: sft
template: default
warmup_steps: 10