badam_mode: layer
badam_switch_interval: 7
badam_switch_mode: ascending
badam_update_ratio: 0.05
bf16: true
cutoff_len: 10240
dataset: api_eval_dataset
dataset_dir: data
ddp_timeout: 180000000
do_train: true
finetuning_type: full
flash_attn: fa2
gradient_accumulation_steps: 4
include_num_input_tokens_seen: true
learning_rate: 3.0e-06
logging_steps: 5
lr_scheduler_type: cosine
max_grad_norm: 1.0
max_samples: 10000000
model_name_or_path: microsoft/Phi-3.5-mini-instruct
num_train_epochs: 32.0
optim: adamw_torch
output_dir: saves/Custom/full/SpecAI
packing: false
per_device_train_batch_size: 4
plot_loss: true
preprocessing_num_workers: 16
report_to: none
save_steps: 2000
shift_attn: true
stage: sft
template: phi
use_badam: true
warmup_steps: 0