dataset_name: mnoukhov/openai_summarize_comparisons_relabel_pythia1b
model_name: /home/toolkit/huggingface/tldr_sft_pythia160m_fp32
load_in_8bit: False
bf16: False
fp16: True
per_device_train_batch_size: 16
per_device_eval_batch_size: 16
gradient_accumulation_steps: 2
gradient_checkpointing: False
learning_rate: 1e-5
optimizer_type: adamw_torch
num_train_epochs: 3
use_lora: True
lora_all_linear: True
lora_r: 8
lora_alpha: 32