5CD-AI/Vintern-3B-beta

Vintern-3B-beta ❄️ - The LLaVA 🌋 Challenger

What's new in Vintern-3B-beta!

• We successfully reproduced the training process of InternVL from scratch.
• The model is the result of integrating Qwen/Qwen2.5-3B-Instruct and InternViT-300M-448px through an MLP layer.
• Trained with more than 10 Milion Vietnamese QnAs, Descriptions, and 10% English Data from OpenGVLab/InternVL-Chat-V1-2-SFT-Data.

Model Details

Model Name	Vision Part	Language Part
Vintern-3B-beta	InternViT-300M-448px	Qwen2.5-3B-Instruct

Bytedance/MTVQA Benchmark

We surpassed GPT-4o and are approaching Gemini 1.5 Pro on the MTVQA dataset for Vietnamese. The benchmark result in MTVQA from open_vlm_leaderboard.

Rank	Method	Param (B)	Language Model	Vision Model	VI
1	Gemini-1.5-Pro				41.3
2	Vintern-3B-beta	3	Qwen2.5-3B-Instruct	InternViT-300M	41.289
3	GPT-4o (0513, detail-h...)				39.6
4	GPT-4o (0806, detail-h...)				38.9
5	Gemini-1.5-Flash				38.9
6	Qwen-VL-Max-0809	72	Qwen2-72B	ViT-600M	36.9
7	GPT-4o (0513, detail-lo...)				26.1
8	Qwen-VL-Plus-0809				27.8
9	GLM-4v-9B	9	GLM-4-9B	EVA-02-5B	26.6
10	InternVL2-Llama3-76B	76	Llama-3-70B-Instruct	InternViT-6B	26.7
11	Step-1.5V		Step-1.5	stepencoder	18.4
12	InternVL2-40B	40	Nous-Hermes-2-Yi-34B	InternViT-6B	21.2
13	Pixtral-12B	13	Nemo-12B	ViT-400M	19.7

Zalo VMLU Benchmark

The Vintern-3B-beta achieved a score of 54.81 on the Zalo VMLU Benchmark.

generation_config = dict(max_new_tokens= 64, do_sample=False, num_beams = 1, repetition_penalty=1.5)
question = "Bạn là trợ lý AI giải trắc nghiệm rất chính xác. Bạn biết chắc chắn đáp án đúng nhất. Chỉ đưa ra chữ cái đứng trước câu trả lời đúng của câu hỏi trắc nghiệm sau: Các cơ quan nào sau đây là cơ quan tư pháp? Lựa Chọn:\nA. Viện kiểm sát nhân dân\nB. Tòa án nhân dân\nC. Chính phủ\nD. Cả A và B\nCâu trả lời đúng nhất là:"
model.chat(tokenizer, None, question, generation_config)

OpenCompass Benchmark

We are creating a pull request for the OpenCompass team to test once more and make the metrics public on the open_vlm_leaderboard.

The current results are at a quite good level, and we are expanding the training set in English and other languages to approach models within a comparable parameter range.

"The table is referenced from the repo Qwen/Qwen2-VL-2B-Instruct."

Benchmark	InternVL2-2B	MiniCPM-V 2.0	Qwen2-VL-2B	Vintern-3B-beta
MMMUval	36.3	38.2	41.1	43.55
DocVQAtest	86.9	-	90.1	80.47
InfoVQAtest	58.9	-	65.5	48.28
ChartQAtest	76.2	-	73.5	68.32
TextVQAval	73.4	-	79.7	67.09
OCRBench	781	605	794	619
MTVQA	10.9	8.8	20.0	23.58
Vi-MTVQA	9.3	8.4	-	41.29
RealWorldQA	57.3	55.8	62.9	57.9
MMEsum	1876.8	1808.6	1872.0	1772.9
MMBench-ENtest	73.2	69.1	74.9	70.62
MMStar	49.8	39.1	48.0	47.6
HallBenchavg	38.0	36.1	41.7	43.22
MathVistatestmini	46.0	39.8	43.0	43.9

Quickstart

Here provides a code snippet to show you how to load the tokenizer and model and how to generate contents. To run inference using the model, follow the steps outlined in our Colab inference notebook

import numpy as np
import torch
import torchvision.transforms as T
# from decord import VideoReader, cpu
from PIL import Image
from torchvision.transforms.functional import InterpolationMode
from transformers import AutoModel, AutoTokenizer

IMAGENET_MEAN = (0.485, 0.456, 0.406)
IMAGENET_STD = (0.229, 0.224, 0.225)

def build_transform(input_size):
    MEAN, STD = IMAGENET_MEAN, IMAGENET_STD
    transform = T.Compose([
        T.Lambda(lambda img: img.convert('RGB') if img.mode != 'RGB' else img),
        T.Resize((input_size, input_size), interpolation=InterpolationMode.BICUBIC),
        T.ToTensor(),
        T.Normalize(mean=MEAN, std=STD)
    ])
    return transform

def find_closest_aspect_ratio(aspect_ratio, target_ratios, width, height, image_size):
    best_ratio_diff = float('inf')
    best_ratio = (1, 1)
    area = width * height
    for ratio in target_ratios:
        target_aspect_ratio = ratio[0] / ratio[1]
        ratio_diff = abs(aspect_ratio - target_aspect_ratio)
        if ratio_diff < best_ratio_diff:
            best_ratio_diff = ratio_diff
            best_ratio = ratio
        elif ratio_diff == best_ratio_diff:
            if area > 0.5 * image_size * image_size * ratio[0] * ratio[1]:
                best_ratio = ratio
    return best_ratio

def dynamic_preprocess(image, min_num=1, max_num=12, image_size=448, use_thumbnail=False):
    orig_width, orig_height = image.size
    aspect_ratio = orig_width / orig_height

    # calculate the existing image aspect ratio
    target_ratios = set(
        (i, j) for n in range(min_num, max_num + 1) for i in range(1, n + 1) for j in range(1, n + 1) if
        i * j <= max_num and i * j >= min_num)
    target_ratios = sorted(target_ratios, key=lambda x: x[0] * x[1])

    # find the closest aspect ratio to the target
    target_aspect_ratio = find_closest_aspect_ratio(
        aspect_ratio, target_ratios, orig_width, orig_height, image_size)

    # calculate the target width and height
    target_width = image_size * target_aspect_ratio[0]
    target_height = image_size * target_aspect_ratio[1]
    blocks = target_aspect_ratio[0] * target_aspect_ratio[1]

    # resize the image
    resized_img = image.resize((target_width, target_height))
    processed_images = []
    for i in range(blocks):
        box = (
            (i % (target_width // image_size)) * image_size,
            (i // (target_width // image_size)) * image_size,
            ((i % (target_width // image_size)) + 1) * image_size,
            ((i // (target_width // image_size)) + 1) * image_size
        )
        # split the image
        split_img = resized_img.crop(box)
        processed_images.append(split_img)
    assert len(processed_images) == blocks
    if use_thumbnail and len(processed_images) != 1:
        thumbnail_img = image.resize((image_size, image_size))
        processed_images.append(thumbnail_img)
    return processed_images

def load_image(image_file, input_size=448, max_num=12):
    image = Image.open(image_file).convert('RGB')
    transform = build_transform(input_size=input_size)
    images = dynamic_preprocess(image, image_size=input_size, use_thumbnail=True, max_num=max_num)
    pixel_values = [transform(image) for image in images]
    pixel_values = torch.stack(pixel_values)
    return pixel_values

model = AutoModel.from_pretrained(
    "5CD-AI/Vintern-3B-beta",
    torch_dtype=torch.bfloat16,
    low_cpu_mem_usage=True,
    trust_remote_code=True,
).eval().cuda()
tokenizer = AutoTokenizer.from_pretrained("5CD-AI/Vintern-3B-beta", trust_remote_code=True, use_fast=False)

test_image = 'test-image.jpg'

pixel_values = load_image(test_image, max_num=6).to(torch.bfloat16).cuda()
generation_config = dict(max_new_tokens= 512, do_sample=False, num_beams = 3, repetition_penalty=3.5)

question = '<image>\nMô tả hình ảnh một cách chi tiết.'

response, history = model.chat(tokenizer, pixel_values, question, generation_config, history=None, return_history=True)
print(f'User: {question}\nAssistant: {response}')

#question = "Câu hỏi khác ......"
#response, history = model.chat(tokenizer, pixel_values, question, generation_config, history=history, return_history=True)
#print(f'User: {question}\nAssistant: {response}')

Citation

@misc{doan2024vintern1befficientmultimodallarge,
      title={Vintern-1B: An Efficient Multimodal Large Language Model for Vietnamese}, 
      author={Khang T. Doan and Bao G. Huynh and Dung T. Hoang and Thuc D. Pham and Nhat H. Pham and Quan T. M. Nguyen and Bang Q. Vo and Suong N. Hoang},
      year={2024},
      eprint={2408.12480},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2408.12480}, 
}