test / llama3_guardrail_lora_sft.yaml
minhbui's picture
Upload 2 files
4324f66 verified
raw
history blame contribute delete
730 Bytes
### examples/train_lora/llama3_lora_sft.yaml
model_name_or_path: "meta-llama/Llama-Guard-3-1B"
stage: sft
do_train: true
do_eval: true
finetuning_type: lora
lora_target: all
dataset: train_sample
eval_dataset: test_sample
dataset_dir: ./data
template: llama3
cutoff_len: 4096
max_samples: 1000
overwrite_cache: true
preprocessing_num_workers: 64
output_dir: ./saves/llama3-1b/lora/sft
logging_steps: 1
save_steps: 10
plot_loss: true
overwrite_output_dir: true
per_device_train_batch_size: 4
gradient_accumulation_steps: 8
learning_rate: 1.0e-4
num_train_epochs: 3.0
lr_scheduler_type: cosine
warmup_ratio: 0.1
bf16: true
ddp_timeout: 180000000
# val_size: 0.1
per_device_eval_batch_size: 16
eval_strategy: steps
eval_steps: 1