33 lines
770 B
YAML
33 lines
770 B
YAML
|
## 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
|
||
|
|
max_target_length: 169
|
||
|
|
## 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: 0.5
|
||
|
|
max_steps: -1
|
||
|
|
num_train_epochs: 1
|
||
|
|
load_in_8bit: False
|
||
|
|
bf16: False
|
||
|
|
fp16: True
|
||
|
|
learning_rate: 1e-6
|
||
|
|
lr_scheduler_type: constant_with_warmup
|
||
|
|
use_peft: True
|
||
|
|
lora_all_linear: True
|
||
|
|
lora_r: 32
|
||
|
|
lora_alpha: 64
|
||
|
|
lora_dropout: 0.05
|
||
|
|
gradient_accumulation_steps: 4
|
||
|
|
per_device_train_batch_size: 4
|
||
|
|
per_device_eval_batch_size: 4
|
||
|
|
warmup_steps: 150
|