We introduce Vintern-1B-v3.5, the latest version in the Vintern series, offering significant improvements over v2 across all evaluation benchmarks. This model has been fine-tuned from InternVL2.5-1B, which already good in Vietnamese 🇻🇳 tasks because it used Viet-ShareGPT-4o-Text-VQA data during its fine-tuning process by the InternVL 2.5 [1] team.

To further enhance its performance in Vietnamese while maintaining good capabilities on existing English datasets, Vintern-1B-v3.5 has been fine-tuned using a vast amount of Vietnamese-specific data. This results in a model that is exceptionally powerful in text recognition, OCR, and understanding Vietnam-specific documents.

Highlights 🌟

Top Quality for Vietnamese Texts Vintern-1B-v3.5 is one of the best models in its class (1B parameters) for understanding and processing Vietnamese documents.
Better Extraction and Understanding The model is great at handling invoices, legal texts, handwriting, and tables.
Improved Prompt Understanding It can understand more complex prompts compared to v2, making it easier to work with.
Runs on Affordable Hardware You can run the model on Google Colab with a T4 GPU, making it easy to use without expensive devices.
Easy to Fine-tune The model can be customized for specific tasks with minimal effort.

🤗 HF Demo 🤗

Benchmarks 📈

Benchmark	InternVL2_5 1B	Vintern-1B-v2	Vintern-1B-v3.5
vi-MTVQA	24.8	37.4	41.9
DocVQAtest	84.8	72.5	78.8
InfoVQAtest	56.0	38.9	46.4
TextVQAval	72.0	64.0	68.2
ChartQAtest	75.9	34.1	65.7
OCRBench	785	628	706

Llama.cpp - Realtime CPU demo by ngxson.

Github: https://github.com/ngxson/vintern-realtime-demo Thanks Son :)

Examples

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

Open In Colab

python

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-1B-v3_5",
    torch_dtype=torch.bfloat16,
    low_cpu_mem_usage=True,
    trust_remote_code=True,
    use_flash_attn=False,
).eval().cuda()

tokenizer = AutoTokenizer.from_pretrained("5CD-AI/Vintern-1B-v3_5", 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= 1024, do_sample=False, num_beams = 3, repetition_penalty=2.5)

question = '<image>\nTrích xuất thông tin chính trong ảnh và trả về dạng markdown.'

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}')

🚀🛠️ Fine-tune using your own data 📂✨

https://colab.research.google.com/drive/1Iu6ssNvlA2bNGlOLq7xR0l-5qtBIqul4

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}, 
}

Reference

[1] Z. Chen et al., ‘Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling’, arXiv preprint arXiv:2412. 05271, 2024.