| export TRANSFORMERS_CACHE=/fs/gamma-projects/audio/gama/hf_cache/ | |
| export HF_DATASETS_CACHE=/fs/gamma-projects/audio/gama/hf_cache/ | |
| output_dir='/fs/gamma-projects/audio/gama/new_data/stage5_all_mix_all_new' | |
| mkdir -p $output_dir | |
| cp "$0" ${output_dir}/$(date +"%Y-%m-%d-%H-%M-%S").sh | |
| torchrun --nproc_per_node=4 --master_port=1234 ../finetune.py \ | |
| --base_model '/fs/gamma-projects/audio/gama/new_data/stage4_all_mix_new/checkpoint-46800/pytorch_model.bin' \ | |
| --data_path '/fs/nexus-projects/brain_project/icml/ltu_all_nexus.json' \ | |
| --output_dir $output_dir \ | |
| --batch_size 164 \ | |
| --micro_batch_size 2 \ | |
| --num_epochs 1 \ | |
| --learning_rate 1e-4 \ | |
| --cutoff_len 108 \ | |
| --val_set_size 0 \ | |
| --lora_r 8 \ | |
| --lora_alpha 16 \ | |
| --lora_dropout 0.05 \ | |
| --lora_target_modules '[q_proj,v_proj]' \ | |
| --train_on_inputs \ | |
| --group_by_length \ | |
| --wandb_run_name ${output_dir} \ | |
| --save_steps 100 \ | |
| --trainable_params qformer_all |