Skip to content

train

Directory actions

More options

Directory actions

More options

Latest commit

 

History

History

train

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
ms-swift-replace-code/swift/llm
ms-swift-replace-code/swift/llm
 
 
qwen_vl_utils
qwen_vl_utils
 
 
scripts
scripts
 
 
README.md
README.md
 
 
requirements.txt
requirements.txt
 
 

README.md

SFT phase

We use ms-swift with some modifications in the SFT phase to train MMDuet2. When loading model (Qwen2.5-VL) and data, we use qwen-vl-utils and transformers.

Training Process

Create conda environment

conda create --name mmduet2_sft python=3.10 conda activate mmduet2_sft pip install -r requirements.txt

Download and replace some code files ms-swift

git clone https://github.com/modelscope/ms-swift.git cd ms-swift git checkout v3.2.0 pip install -e . cd .. # replace some code files that we modified from ms-swift cp -ri ms-swift-replace-code/* ms-swift

Download training data

Follow the instructions in MMDuet2-data to prepare the dataset, and move the sft sub folder to ./data/annotations

Train!

bash ./scripts/train.sh