mode: eval
push_to_hub: False
gold_eval: none
## costa stuff
model_name: mnoukhov/EleutherAI_pythia-1b-deduped__sft__tldr_dpo_1b_fp16.yml_24e9f83_merged
model_revision: step2324
dataset_name: vwxyzjn/summarize_from_feedback_oai_preprocessing_1706381144
tokenizer_name: EleutherAI/pythia-1b-deduped
prompt_field: query
eval_split: validation
max_target_length: 128
## hub stuff
push_to_hub_organization: mnoukhov
## training stuff
eval_steps: 0.2
save_steps: 0.2
beta: 0.5
max_steps: -1
num_train_epochs: 2
load_in_8bit: False
bf16: False
fp16: True
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: 4
per_device_eval_batch_size: 4
warmup_steps: 150