bf16: true
cutoff_len: 512
dataset: /p/data1/mmlaion/marianna/lf_datasets/tatsu-lab/alpaca
dataset_dir: ONLINE
ddp_timeout: 180000000
deepspeed: dcft/train/zero1.json
do_train: true
enable_liger_kernel: true
eval_strategy: epoch
finetuning_type: full
formatting: alpaca
gradient_accumulation_steps: 4
gradient_checkpointing: true
hub_model_id: mlfoundations-dev/tinyllama_alpaca_sft_sample
learning_rate: 2.0e-05
logging_steps: 10
lr_scheduler_type: cosine
model_name_or_path: TinyLlama/TinyLlama-1.1B-Chat-v1.0
neat_packing: true
num_train_epochs: 1.0
output_dir: /p/data1/mmlaion/marianna/dcft_checkpoints/experiments/train/checkpoints/tinyllama_alpaca_sft_sample
overwrite_cache: true
overwrite_output_dir: true
packing: true
per_device_train_batch_size: 8
plot_loss: true
preprocessing_num_workers: 16
push_to_db: false
push_to_hub: false
report_to: wandb
run_name: tinyllama_alpaca_sft_sample
save_strategy: epoch
stage: sft
template: alpaca
val_size: 0.05
warmup_ratio: 0.1