Llama-3.2-MAAL-11B-Vision-v0.1

Llama-3.2-MAAL-11B-Vision-v0.1 is bilingual multimodal model trained for text and visual understanding across Korean and English languages. We are releasing a model, a subset of the training dataset, and a leaderboard to promote and accelerate the development of Korean Vision-Language Models (VLMs).

• Developed by: maum.ai Brain NLP. Jaeyoon Jung, Yoonshik Kim, Yekyung Nah
• Language(s) (NLP): Korean, English (currently, bilingual)

Model Description

Version 0.1 is fine-tuned by English and Korean VQA datasets with other datasets (OCR, Math, etc)...

• We trained this model on 8 H100-80G for 2 days with image-text pair multimodal fine-tuning dataset
• maum-ai/General-Evol-VQA is one of the datasets that we used for fine-tuning.

sample inference code (GPU)

Starting with transformers >= 4.45.0 onward, you can run inference to generate text based on an image and a starting prompt you supply.

import requests
import torch
from PIL import Image
from transformers import MllamaForConditionalGeneration, AutoProcessor

model_id = "maum-ai/Llama-3.2-MAAL-11B-Vision-v0.1"

model = MllamaForConditionalGeneration.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
processor = AutoProcessor.from_pretrained(model_id)

url = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/0052a70beed5bf71b92610a43a52df6d286cd5f3/diffusers/rabbit.jpg"
image = Image.open(requests.get(url, stream=True).raw)

messages = [
    {"role": "user", "content": [
        {"type": "image"},
        {"type": "text", "text": "이 이미지에 대해서 시를 써줘"}
    ]}
]
input_text = processor.apply_chat_template(messages, add_generation_prompt=True)
inputs = processor(
    image,
    input_text,
    add_special_tokens=False,
    return_tensors="pt"
).to(model.device)

output = model.generate(**inputs, max_new_tokens=200)
print(processor.decode(output[0]))

Evaluation Results

As the main goal of version 0.1 is leveraging Korean VQA and OCR capabilities tailored to real-world business use cases, we select KOFFVQA as our evaluation method to assess the Korean instruction-following skills.

Model	Params (B)	average(↑)
NCSOFT/VARCO-VISION-14B	15.2b	66.69
Qwen/Qwen2-VL-7B-Instruct	8.3b	63.53
maum-ai/Llama-3.2-MAAL-11B-Vision-v0.1	10.7b	61.13
meta-llama/Llama-3.2-11B-Vision-Instruct	10.7b	50.36
mistralai/Pixtral-12B-2409	12.7b	44.62
llava-onevision-qwen2-7b-ov	8b	43.78
InternVL2-8b	8.1b	32.76
MiniCPM-V-2_6	8.1b	32.69

Our model has achieved a 20% performance improvement compared to the previous base model. You can check more results in this Leaderboard

We will release enhanced model, v0.2 soon