# TODO:
- [x] Save LoRA separately
- [ ] Load LoRA separately
- [ ] Merge LoRA

# How to use
1. Setup the environment
```bash
conda create -n llava python=3.10 -y
conda activate llava
pip install --upgrade pip  # Enable PEP 660 support.
pip install -e ".[train]"
```

2. Dataset
The dataset is stored in a json file. Each item is in the following format:
```json
{
        "id": "<id>",
        "image": "/path/to/image",
        "conversations": [
            {
                "from": "human",
                "value": "<question>"
            },
            {
                "from": "gpt",
                "value": "<answer>"
            }
        ],
        ...
}
```
Modify the `--data_path` flag, which should be a folder containing `train.json` and `test.json`.

3. Pre-trained ckpt
Usually, huggingface will download the pretrained ckpt automatically. But sometimes it will take a lot of time if the server is in mainland. Alternatively, you can find the pretrained ckpt under `61.170.32.4:/mnt1/lyc/huggingface/hub`.

4. Finetuning
Sample bash
```bash
bash scripts/finetune/train_llava.sh
```


5. Testing
Sample bash
```bash
bash scripts/finetune/test_llava.sh
```