model_name: /home/toolkit/huggingface/tldr_sft_pythia7b_4V100_seq550
load_in_8bit: False
bf16: False
fp16: True
per_device_train_batch_size: 2
per_device_eval_batch_size: 2
gradient_accumulation_steps: 16
gradient_checkpointing: False
learning_rate: 1e-5
optimizer_type: paged_adamw_32bit
num_train_epochs: 5
seq_length: 560