SFT_GUIDE.md

Instruction Fine-Tuning Guide

Data Downloading

The data will be uploaded to 🤗 VLM-SFT soon.

# assume you have installed git-lfs. If not, please run conda install git-lfs.
git clone https://huggingface.co/datasets/YangyiYY/VLM-SFT

Model Training

You should first specify the train_data, img_dir, proj_dir, checkpoint in the config/SFT.yml file:

1. train_data: the path to the training data all_data.jsonl (downloaded in the previous step).
2. img_dir: the path to the image directory. The images are downloaded in the previous step.
3. proj_dir: the name for wandb logger. You can set it to your own project name.
4. checkpoint: the path to the pre-trained model. See PRETRAIN_GUIDE.md for the pre-training instructions.

Then run:

scripts/sft/run.sh

The output will be saved to data/ckpts/SFT/.