bf16: true
cutoff_len: 1024
dataset: MAIR-GEN-mini-v2-50000
dataset_dir: data
ddp_timeout: 180000000
deepspeed: cache/ds_z2_config.json
do_train: true
finetuning_type: full
flash_attn: auto
gradient_accumulation_steps: 4
include_num_input_tokens_seen: true
learning_rate: 5.0e-05
logging_steps: 5
lr_scheduler_type: cosine
max_grad_norm: 1.0
max_samples: 10000000
model_name_or_path: meta-llama/Llama-3.2-3B-Instruct
num_train_epochs: 3.0
optim: adamw_torch
output_dir: /usr1/data/weiweis/QG/models/train_2025-03-19-00-21-09
packing: false
per_device_train_batch_size: 4
plot_loss: true
preprocessing_num_workers: 16
report_to: none
save_steps: 5000
stage: sft
template: llama3
train_on_prompt: true
trust_remote_code: true
warmup_steps: 200