model_name: mnoukhov/pythia410m-tldr-sft
dataset_name: mnoukhov/openai_summarize_comparisons_tldrprompt_relabel1b
train_split: train[:1]
gold_model_name: mnoukhov/pythia1b-sft-rm-tldrprompt
pseudo_dataset_name: mnoukhov/openai_summarize_generated_20k_relabel_410m_dpo1
beta: 0.5
max_steps: 10000
eval_steps: 1000
load_in_8bit: False
bf16: True
fp16: False
learning_rate: 1e-5
use_peft: True
lora_all_linear: True
lora_r: 8
lora_alpha: 32
lora_dropout: 0.05
gradient_accumulation_steps: 4
per_device_train_batch_size: 16
warmup_steps: 150