## dpo 2 
pseudo_dataset_name: mnoukhov/summarize_from_feedback_tldr3_generated_relabel_20k_dpo_costa_1b_fp16.yml_3d94f50_b9ff2
train_split: train[:1]
max_prompt_length: 512
max_target_length: 131
max_length: 640
lr_scheduler_type: cosine
## costa stuff
model_name: vwxyzjn/EleutherAI_pythia-1b-deduped__sft__tldr
model_revision: sft__55513__1706646024
dataset_name: vwxyzjn/summarize_from_feedback_oai_preprocessing_1706381144
tokenizer_name: EleutherAI/pythia-1b-deduped
prompt_field: query
eval_split: validation
## hub stuff
push_to_hub: True
push_to_hub_organization: mnoukhov
## training stuff
gold_eval: ppl
eval_steps: 0.2
save_steps: 0.2
beta: 0.05
max_steps: -1
num_train_epochs: 2
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