gemma3_chosen_lora / training_args.yaml
AndyPark's picture
Upload folder using huggingface_hub
a6509a7 verified
bf16: true
cutoff_len: 9000
dataset: chosen_22607_250423
dataset_dir: data
ddp_timeout: 180000000
do_train: true
double_quantization: true
finetuning_type: lora
flash_attn: auto
gradient_accumulation_steps: 8
include_num_input_tokens_seen: true
learning_rate: 5.0e-05
logging_steps: 5
lora_alpha: 16
lora_dropout: 0.1
lora_rank: 8
lora_target: all
lr_scheduler_type: constant_with_warmup
max_grad_norm: 1.0
max_samples: 100000
model_name_or_path: /workspace/LLaMA-Factory/models/gemma-3-27b-it
num_train_epochs: 2.0
optim: adamw_torch
output_dir: saves/Gemma-3-27B-Instruct/lora/DPO_chosen_22607_250423
packing: false
per_device_train_batch_size: 1
plot_loss: true
pref_beta: 0.1
pref_ftx: 0
pref_loss: sigmoid
preprocessing_num_workers: 16
quantization_bit: 4
quantization_method: bnb
report_to: none
save_steps: 1000
stage: dpo
template: gemma3
trust_remote_code: true
warmup_steps: 100