### model
model_config_path: model_configs/model_config_small.json
tokenizer: data/tokenizer

### dataset
data_path: data/raw_dataset/VirusComment.py
dataloader_num_workers: 8

### output
run_name: viruscommentcls-training
output_dir: data/output_models/viruscommentcls_12
overwrite_output_dir: true
save_strategy: epoch
report_to: tensorboard
logging_steps: 1

### train
do_train: true
num_train_epochs: 25
per_device_train_batch_size: 32
gradient_accumulation_steps: 32
optim: adamw_torch
learning_rate: 4.3e-5
lr_scheduler_type: cosine
warmup_ratio: 0.05
fp16: true
ddp_backend: null
ddp_timeout: 180000000
loss_type: 'focal'
focal_alpha: 1.0
focal_gamma: 2.0
label_smoothing: 0.1

### eval
do_eval: true
per_device_eval_batch_size: 128
eval_strategy: steps
eval_steps: 0.02