training_args.yaml

bf16: true
cutoff_len: 2048
dataset: followir_advanced_1
dataset_dir: data
ddp_timeout: 180000000
do_train: true
finetuning_type: lora
flash_attn: auto
gradient_accumulation_steps: 32
learning_rate: 3.0e-05
logging_steps: 5
lora_alpha: 16
lora_dropout: 0
lora_rank: 8
lora_target: q_proj,v_proj,o_proj,k_proj
lr_scheduler_type: cosine
max_grad_norm: 1.0
max_samples: 100000
model_name_or_path: mistralai/Mistral-7B-Instruct-v0.2
num_train_epochs: 8.0
optim: adamw_torch
output_dir: saves/Mistral-7B-Instruct-v0.2/lora/train_2025-01-21-12-40-17
packing: false
per_device_train_batch_size: 4
plot_loss: true
preprocessing_num_workers: 16
report_to: none
save_steps: 100
stage: sft
template: mistral
trust_remote_code: true
warmup_steps: 0