[Question] loss is too low, only 0.3

[an1018]

Question

Training with custom data (12,276 images, 30 images from llava158K), after 47iters loss dropped to 0.4. Training script: deepspeed --include localhost:5,6 --master_port 29585 \ llava/train/train_mem.py \ --lora_enable True --lora_r 128 --lora_alpha 256 --mm_projector_lr 2e-5 \ --deepspeed ./scripts/zero3.json \ --model_name_or_path ./llava-v1.5-13b \ --version v1 \ --data_path ../dataset/llava_finetune/train.json \ --image_folder ../dataset/llava_finetune/ \ --vision_tower openai/clip-vit-large-patch14-336 \ --mm_projector_type mlp2x_gelu \ --mm_vision_select_layer -2 \ --mm_use_im_start_end False \ --mm_use_im_patch_token False \ --image_aspect_ratio pad \ --group_by_modality_length True \ --bf16 True \ --output_dir ./checkpoints/llava-v1.5-13b-task-lora-nuimages \ --num_train_epochs 10 \ --per_device_train_batch_size 16 \ --per_device_eval_batch_size 4 \ --gradient_accumulation_steps 1 \ --evaluation_strategy "no" \ --save_strategy "steps" \ --save_steps 384 \ --save_total_limit 11 \ --learning_rate 2e-5 \ --weight_decay 0. \ --warmup_ratio 0.03 \ --lr_scheduler_type "cosine" \ --logging_steps 1 \ --tf32 True \ --model_max_length 2048 \ --gradient_checkpointing True \ --dataloader_num_workers 4 \ --lazy_preprocess True \ --report_to wandb this is my prompt:

I have tried to change the dataset twice, could you give me some advice? Thank you very much

[an1018]

@haotian-liu Could you help me see how I should change it

[Jeckinchen]

any update?

[fisher75]

Hi, how do you know the training was effecitve? Did you use the default training setting? I LoRA with default parameters and basically no improvement.

[bang123-box]

Question

Training with custom data (12,276 images, 30 images from llava158K), after 47iters loss dropped to 0.4. Training script: deepspeed --include localhost:5,6 --master_port 29585 \ llava/train/train_mem.py \ --lora_enable True --lora_r 128 --lora_alpha 256 --mm_projector_lr 2e-5 \ --deepspeed ./scripts/zero3.json \ --model_name_or_path ./llava-v1.5-13b \ --version v1 \ --data_path ../dataset/llava_finetune/train.json \ --image_folder ../dataset/llava_finetune/ \ --vision_tower openai/clip-vit-large-patch14-336 \ --mm_projector_type mlp2x_gelu \ --mm_vision_select_layer -2 \ --mm_use_im_start_end False \ --mm_use_im_patch_token False \ --image_aspect_ratio pad \ --group_by_modality_length True \ --bf16 True \ --output_dir ./checkpoints/llava-v1.5-13b-task-lora-nuimages \ --num_train_epochs 10 \ --per_device_train_batch_size 16 \ --per_device_eval_batch_size 4 \ --gradient_accumulation_steps 1 \ --evaluation_strategy "no" \ --save_strategy "steps" \ --save_steps 384 \ --save_total_limit 11 \ --learning_rate 2e-5 \ --weight_decay 0. \ --warmup_ratio 0.03 \ --lr_scheduler_type "cosine" \ --logging_steps 1 \ --tf32 True \ --model_max_length 2048 \ --gradient_checkpointing True \ --dataloader_num_workers 4 \ --lazy_preprocess True \ --report_to wandb this is my prompt:

I have tried to change the dataset twice, could you give me some advice? Thank you very much

Same as me, I use my custom data to pretrain and lora sft the llava, when the pretrained is done, the loss is around to 0.4-0.5, when sft stage, the loss decreases rapidly from 1.0 to 0.4, then the loss is around to 0.2-04