## model and dataset
model_name: "mnoukhov/pythia-2.8b-mitchell-sft_hh_rlhf"
# output name
hub_model_id: "mnoukhov/pythia-2.8b-mitchell-dpo_hh_rlhf"
dataset_name: "sophiex/hh-rlhf"
report_to: "wandb"
## dpo
learning_rate: 0.0000005
lr_scheduler_type: constant_with_warmup
warmup_steps: 150
fp16: False
bf16: True
gradient_accumulation_steps: 16
per_device_train_batch_size: 4
per_device_eval_batch_size: 4
num_train_epochs: 1
max_length: 512
max_prompt_length: 256 
max_target_length: 256
beta: 0.1
## peft
use_peft: False
# lora_r: 16
# lora_alpha: 32
gradient_checkpointing: True
evaluation_strategy: "steps"
eval_steps: 0.2
logging_steps: 100
ddp_find_unused_parameters: False