# model
model_name_or_path: google/paligemma-3b-mix-448
visual_inputs: true
tune_mm_proj: true
#print_param_status: true

# method
stage: sft
do_train: true
finetuning_type: full

# ddp
ddp_timeout: 180000000
deepspeed: examples/deepspeed/ds_z2_offload_config.json

# dataset
dataset: mllm_pt_demo
dataset_dir: data
template: gemma
cutoff_len: 2048
max_samples: 3
#val_size: 0.0001
overwrite_cache: true
preprocessing_num_workers: 16

# output
output_dir: saves/paligemma/full/sft_llava_pt_test
logging_steps: 1
save_steps: 50
plot_loss: true
overwrite_output_dir: true
#save_strategy: epoch
#save_total_limit: 2

# train
per_device_train_batch_size: 1
gradient_accumulation_steps: 16
learning_rate: 0.00001
num_train_epochs: 100
lr_scheduler_type: cosine
warmup_steps: 0.1
#bf16: true
pure_bf16: true

# eval
do_eval: false
#per_device_eval_batch_size: 1
#evaluation_strategy: steps
#eval_steps: 500