## costa stuff
model_name: sophiex/pythia-2.8b-sft_hh_rlhf
# model_revision: null
dataset_name: sophiex/hh-rlhf
tokenizer_name: EleutherAI/pythia-2.8b-deduped
prompt_field: prompt
eval_split: test
max_prompt_length: 256
max_target_length: 256
max_length: 512
lr_scheduler_type: cosine
## hub stuff
push_to_hub: True
push_to_hub_organization: mnoukhov
## training stuff
save_strategy: steps
gold_eval: ppl
gold_dataset_name: sophiex/hh-rlhf
gold_target_field: chosen
gold_eval_split: test
eval_steps: 0.2
save_steps: 0.2
beta: 0.1
max_steps: -1
num_train_epochs: 1
load_in_8bit: False
bf16: False
fp16: True
learning_rate: 1e-5
use_peft: True
lora_r: 16
lora_alpha: 32
lora_dropout: 0.
gradient_accumulation_steps: 4
per_device_train_batch_size: 4
per_device_eval_batch_size: 4