model_name: "EleutherAI/pythia-2.8b-deduped"
output_model_name: "mnoukhov/pythia-2.8b-sft_hh_rlhf"
dataset_name: "sophiex/hh-rlhf"
dataset_text_field: null
report_to: "wandb"
learning_rate: 1e-5
lr_scheduler_type: cosine
fp16: False
bf16: True
gradient_accumulation_steps: 32
per_device_train_batch_size: 4
per_device_eval_batch_size: 4
num_train_epochs: 1
max_seq_length: 1024
use_peft: True
lora_r: 16
lora_alpha: 32
gradient_checkpointing: True
evaluation_strategy: "steps"
eval_steps: 0.2
logging_steps: 100
# ddp_find_unused_parameters: False
sanity_check: True