How to finetune Osprey on RefCOCOg?

[Glupayy]

Hi! Thanks for the great work!

Could you share any configs on fine-tuning Osprey on RefCOCOg dataset? I am trying to follow your work and reproduce the results on it, what's the starting checkpoint and the prompt template? It would be very appreciated if any fine-tuning config could be shared.
Thank you!

[CircleRadon]

Hi, @Glupapa
The starting checkpoint is our final model Osprey-7b, and the prompt template is the same as RefCOCO as in https://github.com/CircleRadon/Osprey/blob/ca9f26dbd9a0907d8ff686588a394fa897b60828/osprey/datasets/stage2_data.py#L256C26-L262C1
The config is as follows,

#!/bin/bash
export PYTHONPATH=`pwd`:$PYTHONPATH

deepspeed --include localhost:0,1,2,3 llava/train/train_mem.py \
    --deepspeed ./scripts/zero2.json \
    --model_name_or_path ./Osprey-7b \
    --dataset_config ./osprey/configs/finetune.json \
    --version v1 \
    --vision_tower laion2b_s29b_b131k_ft_soup.bin \
    --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 './exp/finetune' \
    --num_train_epochs 1 \
    --per_device_train_batch_size 1\
    --per_device_eval_batch_size 4 \
    --gradient_accumulation_steps 1 \
    --evaluation_strategy "no" \
    --save_strategy "steps" \
    --save_steps 3000 \
    --save_total_limit 1 \
    --learning_rate 5e-6 \
    --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 "none" \
    --group_by_modality_length False