---
language:
- ja
pipeline_tag: image-text-to-text
---
# LLM-jp-3 VILA 14B

This repository provides a large vision language model (VLM) developed by the [Research and Development Center for Large Language Models](https://llmc.nii.ac.jp/) at the [National Institute of Informatics](https://www.nii.ac.jp/en/), Japan.

## Usage

Python version: 3.10.12

1. Clone the repository and install the libraries.

    <details>

    ```bash
    git clone git@github.com:llm-jp/llm-jp-VILA.git
    cd llm-jp-VILA
    ```

    ```bash
    python3 -m venv venv
    source venv/bin/activate
    ```

    ```bash
    pip install --upgrade pip
    wget https://github.com/Dao-AILab/flash-attention/releases/download/v2.4.2/flash_attn-2.4.2+cu118torch2.0cxx11abiFALSE-cp310-cp310-linux_x86_64.whl
    pip install flash_attn-2.4.2+cu118torch2.0cxx11abiFALSE-cp310-cp310-linux_x86_64.whl
    pip install -e .
    pip install -e ".[train]"
    ```

    ```bash
    pip install git+https://github.com/huggingface/transformers@v4.36.2
    cp -rv ./llava/train/transformers_replace/* ./venv/lib/python3.10/site-packages/transformers/
    ```

    </details>

2. Run the python script. You can change the `image_path` and `query` to your own.

    <details>

    ```python
    import argparse
    from io import BytesIO

    import requests
    import torch
    from PIL import Image

    from llava.constants import IMAGE_TOKEN_INDEX
    from llava.conversation import conv_templates
    from llava.mm_utils import (get_model_name_from_path,
                                process_images, tokenizer_image_token)
    from llava.model.builder import load_pretrained_model
    from llava.utils import disable_torch_init


    def load_image(image_file):
        if image_file.startswith("http") or image_file.startswith("https"):
            response = requests.get(image_file)
            image = Image.open(BytesIO(response.content)).convert("RGB")
        else:
            image = Image.open(image_file).convert("RGB")
        return image


    def load_images(image_files):
        out = []
        for image_file in image_files:
            image = load_image(image_file)
            out.append(image)
        return out


    disable_torch_init()

    model_checkpoint_path = "llm-jp/llm-jp-3-vila-14b"
    model_name = get_model_name_from_path(model_checkpoint_path)
    tokenizer, model, image_processor, context_len = load_pretrained_model(model_checkpoint_path, model_name)

    image_path = "path/to/image"
    image_files = [
        image_path
    ]
    images = load_images(image_files)

    query = "<image>\nこの画像について説明してください。"

    conv_mode = "llmjp_v3"
    conv = conv_templates[conv_mode].copy()
    conv.append_message(conv.roles[0], query)
    conv.append_message(conv.roles[1], None)
    prompt = conv.get_prompt()

    images_tensor = process_images(images, image_processor, model.config).to(model.device, dtype=torch.float16)
    input_ids = tokenizer_image_token(prompt, tokenizer, IMAGE_TOKEN_INDEX, return_tensors="pt").unsqueeze(0).cuda()

    with torch.inference_mode():
        output_ids = model.generate(
            input_ids,
            images=[
                images_tensor,
            ],
            do_sample=False,
            num_beams=1,
            max_new_tokens=256,
            use_cache=True,
        )

    outputs = tokenizer.batch_decode(output_ids, skip_special_tokens=True)[0]
    print(outputs)
    ```

    </details>

## Model Details

|Model components|Model / Architecture|Parameters|
|:---:|:---:|:---:|
|Vision encoder|[siglip-so400m-patch14-384](https://huggingface.co/google/siglip-so400m-patch14-384)|428M|
|Projector|2-layer MLP|32M|
|LLM|[llm-jp-3-13b-instruct](https://huggingface.co/llm-jp/llm-jp-3-13b-instruct)|13B|

## Datasets

The model was trained in three stages.

### Step-0

We used the following data sets to tune the parameters in the projector.

| Language | Dataset | Images|
|:---|:---|---:|
|Japanese|[Japanese image text pairs](https://gitlab.llm-jp.nii.ac.jp/datasets/llm-jp-japanese-image-text-pairs)|558K
|English|[LLaVA-Pretrain](https://huggingface.co/datasets/liuhaotian/LLaVA-Pretrain)|558K

### Step-1

We used the following data sets to tune the parameters in the projector and LLM.
 
| Language | Dataset | Images |
|:---|:---|:---|
|Japanese|[Japanese image text pairs](https://gitlab.llm-jp.nii.ac.jp/datasets/llm-jp-japanese-image-text-pairs)| 6M |
|        |[Japanese interleaved data](https://gitlab.llm-jp.nii.ac.jp/datasets/llm-jp-japanese-interleaved-data)| 6M |
|English |[coyo](https://github.com/kakaobrain/coyo-dataset) (subset) | 6M | 
| |[mmc4-core](https://github.com/allenai/mmc4) (subset) | 6M | 

### Step-2

We used the following data sets to tune the parameters in the projector and LLM.

| Language | Dataset | Images |
|:---|:---|:---|
|Japanese|[llava-instruct-ja](https://huggingface.co/datasets/llm-jp/llava-instruct-ja)| 156K |
|        |[japanese-photos-conv](https://huggingface.co/datasets/llm-jp/japanese-photos-conversation)| 12K |
|        |[ja-vg-vqa](https://huggingface.co/datasets/llm-jp/ja-vg-vqa-conversation)| 99K |
|        |[synthdog-ja](https://huggingface.co/datasets/naver-clova-ix/synthdog-ja) (subset)| 102K |
|English |[LLaVA](https://huggingface.co/datasets/liuhaotian/LLaVA-Instruct-150K) | 158K | 
|        |[VQAv2](https://visualqa.org/) | 53K | 
|        |[GQA](https://cs.stanford.edu/people/dorarad/gqa/index.html) | 46K | 
|        |[OCRVQA](https://ocr-vqa.github.io/) | 80K | 
|        |[TextVQA](https://textvqa.org/dataset/) | 22K | 

## Evaluations
We evaluated our model using [Heron Bench](https://huggingface.co/datasets/turing-motors/Japanese-Heron-Bench), [JA-VLM-Bench-In-the-Wild](https://huggingface.co/datasets/SakanaAI/JA-VLM-Bench-In-the-Wild), and [JA-VG-VQA500](https://huggingface.co/datasets/SakanaAI/JA-VG-VQA-500). 
We used `gpt-4o-2024-05-13` for LLM-as-a-judge.

### Heron Bench

| Models | LLM-as-a-judge score (%) | 
|---|:---:|
| [Japanese InstructBLIP Alpha](https://huggingface.co/stabilityai/japanese-instructblip-alpha) | 14.0 | 
| [Japanese Stable VLM](https://huggingface.co/stabilityai/japanese-stable-vlm) | 24.2 | 
| [Llama-3-EvoVLM-JP-v2](https://huggingface.co/SakanaAI/Llama-3-EvoVLM-JP-v2) | 39.3 | 
| [LLaVA-CALM2-SigLIP](https://huggingface.co/cyberagent/llava-calm2-siglip) | 43.3 | 
| **llm-jp-3-vila-14b (Ours)** | 57.2 | 
| GPT-4o | 87.6 | 

### JA-VLM-Bench-In-the-Wild

| **Models** | ROUGE-L | LLM-as-a-judge score (/5.0) | 
|---|:---:|:---:|
| [Japanese InstructBLIP Alpha](https://huggingface.co/stabilityai/japanese-instructblip-alpha) | 20.8 | 2.42 | 
| [Japanese Stable VLM](https://huggingface.co/stabilityai/japanese-stable-vlm) | 23.3 | 2.47 | 
| [Llama-3-EvoVLM-JP-v2](https://huggingface.co/SakanaAI/Llama-3-EvoVLM-JP-v2) | 41.4 | 2.92 | 
| [LLaVA-CALM2-SigLIP](https://huggingface.co/cyberagent/llava-calm2-siglip) | 47.2 | 3.15 | 
| **llm-jp-3-vila-14b (Ours)** | 52.3 | 3.69 | 
| GPT-4o | 37.6 | 3.85 | 

### JA-VG-VQA-500

| **Models** | ROUGE-L | LLM-as-a-judge score (/5.0) |
|---|:---:|:---:|
| [Japanese InstructBLIP Alpha](https://huggingface.co/stabilityai/japanese-instructblip-alpha) | -- | -- |
| [Japanese Stable VLM](https://huggingface.co/stabilityai/japanese-stable-vlm) | -- | -- |
| [Llama-3-EvoVLM-JP-v2](https://huggingface.co/SakanaAI/Llama-3-EvoVLM-JP-v2) | 23.5 | 2.96 |
| [LLaVA-CALM2-SigLIP](https://huggingface.co/cyberagent/llava-calm2-siglip) | 17.4 | 3.21 |
| **llm-jp-3-vila-14b (Ours)** | 16.2 | 3.62 |
| GPT-4o | 12.1 | 3.58 |

## Risks and Limitations

The model released in this repository is in the early stages of our research and development. It has not been tuned such that model's outputs are aligned with social norms, ethical standards, and the law.

## License

The weights of this model are released under the [Apache License, Version 2.0](https://www.apache.org/licenses/LICENSE-2.0).
In addition, a user of this model must comply with [the OpenAI terms of use](https://openai.com/policies/terms-of-use) because the model used synthetic data generated by OpenAI GPT-4.

## Additional information

Regarding the license of the [synthdog-ja](https://huggingface.co/datasets/naver-clova-ix/synthdog-ja) dataset, there is no explicit license statement in the dataset documentation. While we attempted to contact the main corresponding author of "OCR-free Document Understanding Transformer" for clarification, we received no response.

Based on the following considerations:

1. The [donut-base](https://huggingface.co/naver-clova-ix/donut-base) model trained on this dataset is released under the MIT license
2. The Wikipedia articles used in the dataset are licensed under CC-BY-SA

We have determined that the synthdog-ja dataset is most likely governed by the CC-BY-SA license, and proceeded with training under this assumption.