README.md

metadata

license: apache-2.0
language:
  - vi
  - en
  - zh
base_model:
  - Qwen/Qwen2-VL-2B-Instruct
library_name: transformers
tags:
  - erax
  - multimodal
  - erax-vl-2B
  - insurance
  - ocr
  - vietnamese
  - bcg
pipeline_tag: visual-question-answering
widget:
  - src: images/photo-1-16505057982762025719470.webp
    example_title: Test 1
  - src: images/vt-don-thuoc-f0-7417.jpeg
    example_title: Test 2

EraX-VL-2B-V1.5

Introduction 🎉

Hot on the heels of the popular EraX-VL-7B-V1.0 model, we proudly present EraX-VL-2B-V1.5. This enhanced multimodal model offers robust OCR and VQA capabilities across diverse languages 🌍, with a significant advantage in processing Vietnamese 🇻🇳. The EraX-VL-2B model stands out for its precise recognition capabilities across a range of documents 📝, including medical forms 🩺, invoices 🧾, bills of sale 💳, quotes 📄, and medical records 💊. This functionality is expected to be highly beneficial for hospitals 🏥, clinics 💉, insurance companies 🛡️, and other similar applications 📋. Built on the solid foundation of the Qwen/Qwen2-VL-2B-Instruct[1], which we found to be of high quality and fluent in Vietnamese, EraX-VL-2B has been fine-tuned to enhance its performance. We plan to continue improving and releasing new versions for free, along with sharing performance benchmarks in the near future.

One standing-out feature of EraX-VL-2B-V1.5 is the capability to do multi-turn Q&A with reasonable reasoning capability at its small size of only +2 billions parameters.

NOTA BENE:

• EraX-VL-2B-V1.5 is NOT a typical OCR-only tool likes Tesseract but is a Multimodal LLM-based model. To use it effectively, you may have to twist your prompt carefully depending on your tasks.
• This model was NOT finetuned with medical (X-ray) dataset or car accidences (yet). Stay tune for updated version coming up sometime 2025.

EraX-VL-2B-V1.5 is a young and tiny member of our EraX's LànhGPT collection of LLM models.

• Developed by:
  • Nguyễn Anh Nguyên ([email protected])
  • Nguyễn Hồ Nam (BCG)
  • Phạm Huỳnh Nhật ([email protected])
  • Phạm Đình Thục ([email protected])
• Funded by: Bamboo Capital Group and EraX
• Model type: Multimodal Transformer with over 2B parameters
• Languages (NLP): Primarily Vietnamese with multilingual capabilities
• License: Apache 2.0
• Fine-tuned from: Qwen/Qwen2-VL-2B-Instruct
• Prompt examples: Some popular prompt examples.

Benchmarks 📊

🏆 LeaderBoard

Models	Open-Source	VI-MTVQA
EraX-VL-7B-V1.5 🥇	✅	47.2
Qwen2-VL 72B 🥈	✘	41.6
ViGPT-VL 🥉	✘	39.1
EraX-VL-2B-V1.5	✅	38.2
EraX-VL-7B-V1	✅	37.6
Vintern-1B-V2	✅	37.4
Qwen2-VL 7B	✅	30.0
Claude3 Opus	✘	29.1
GPT-4o mini	✘	29.1
GPT-4V	✘	28.9
Gemini Ultra	✘	28.6
InternVL2 76B	✅	26.9
QwenVL Max	✘	23.5
Claude3 Sonnet	✘	20.8
QwenVL Plus	✘	18.1
MiniCPM-V2.5	✅	15.3

The test code for evaluating models in the paper can be found in: EraX-JS-Company/EraX-MTVQA-Benchmark

API trial 🎉

Please contact [email protected] for API access inquiry.

Examples 🧩

1. OCR - Optical Character Recognition for Multi-Images

Example 01: Citizen identification card

Front View

Back View

Source: Google Support

{
  "Số thẻ":"037094012351"
  "Họ và tên":"TRỊNH QUANG DUY"
  "Ngày sinh":"04/09/1994"
  "Giới tính":"Nam"
  "Quốc tịch":"Việt Nam"
  "Quê quán / Place of origin":"Tân Thành, Kim Sơn, Ninh Bình"
  "Nơi thường trú / Place of residence":"Xóm 6 Tân Thành, Kim Sơn, Ninh Bình"
  "Có giá trị đến":"04/09/2034"
  "Đặc điểm nhân dạng / Personal identification":"seo chấm c:1cm trên đuôi mắt trái"
  "Cục trưởng cục cảnh sát quản lý hành chính về trật tự xã hội":"Nguyễn Quốc Hùng"
  "Ngày cấp":"10/12/2022"
}

Example 01: Identity Card

Front View

Back View

Source: Internet

{
  "Số":"272737384"
  "Họ tên":"PHẠM NHẬT TRƯỜNG"
  "Sinh ngày":"08-08-2000"
  "Nguyên quán":"Tiền Giang"
  "Nơi ĐKHK thường trú":"393, Tân Xuân, Bảo Bình, Cẩm Mỹ, Đồng Nai"
  "Dân tộc":"Kinh"
  "Tôn giáo":"Không"
  "Đặc điểm nhận dạng":"Nốt ruồi c.3,5cm trên sau cánh mũi phải."
  "Ngày cấp":"30 tháng 01 năm 2018"
  "Giám đốc CA":"T.BÌNH ĐỊNH"
}

Example 02: Driver's License

Front View

Back View

Source: Báo Pháp luật

{
  "No.":"400116012313"
  "Fullname":"NGUYỄN VĂN DŨNG"
  "Date_of_birth":"08/06/1979"
  "Nationality":"VIỆT NAM"
  "Address":"X. Quỳnh Hầu, H. Quỳnh Lưu, T. Nghệ An
  Nghệ An, ngày/date 23 tháng/month 04 năm/year 2022"
  "Hang_Class":"FC"
  "Expires":"23/04/2027"
  "Place_of_issue":"Nghệ An"
  "Date_of_issue":"ngày/date 23 tháng/month 04 năm/year 2022"
  "Signer":"Trần Anh Tuấn"
  "Các loại xe được phép":"Ô tô hạng C kéo rơmoóc, đầu kéo kéo sơmi rơmoóc và xe hạng B1, B2, C, FB2 (Motor vehicle of class C with a trailer, semi-trailer truck and vehicles of classes B1, B2, C, FB2)"
  "Mã số":""
}

Example 03: Vehicle Registration Certificate

Source: Báo Vietnamnet

{
  "Tên chủ xe":"NGUYỄN TÔN NHUẬN"
  "Địa chỉ":"KE27 Kp3 P.TTTây Q7"
  "Nhãn hiệu":"HONDA"
  "Số loại":"DYLAN"
  "Màu sơn":"Trắng"
  "Số người được phép chở":"02"
  "Nguồn gốc":"Xe nhập mới"
  "Biển số đăng ký":"59V1-498.89"
  "Đăng ký lần đầu ngày":"08/06/2004"
  "Số máy":"F03E-0057735"
  "Số khung":"5A04F-070410"
  "Dung tích":"152"
  "Quản lý":"TRƯỞNG CA QUẬN"
  "Thượng tá":"Trần Văn Hiểu"
}

Example 04: Birth Certificate

Source: https://congchung247.com.vn

{
    "name": "NGUYỄN NAM PHƯƠNG",
    "gender": "Nữ",
    "date_of_birth": "08/6/2011",
    "place_of_birth": "Bệnh viện Việt - Pháp Hà Nội",
    "nationality": "Việt Nam",
    "father_name": "Nguyễn Ninh Hồng Quang",
    "father_dob": "1980",
    "father_address": "309 nhà E2 Bạch Khoa - Hai Bà Trưng - Hà Nội",
    "mother_name": "Phạm Thùy Trang",
    "mother_dob": "1984",
    "mother_address": "309 nhà E2 Bạch Khoa - Hai Bà Trưng - Hà Nội",
    "registration_place": "UBND phường Bạch Khoa - Quận Hai Bà Trưng - Hà Nội",
    "registration_date": "05/8/2011",
    "registration_ralation": "cha",
    "notes": None,
    "certified_by": "Nguyễn Thị Kim Hoa"
}

Quickstart 🎮

Install the necessary packages:

python -m pip install git+https://github.com/huggingface/transformers accelerate
python -m pip install qwen-vl-utils
pip install flash-attn --no-build-isolation

Then you can use EraX-VL-2B-V1.5 like this:

import os
import base64
import json

import cv2
import numpy as np
import matplotlib.pyplot as  plt

import torch
from transformers import Qwen2VLForConditionalGeneration, AutoTokenizer, AutoProcessor
from qwen_vl_utils import process_vision_info

model_path = "erax/EraX-VL-2B-V1.5"

model = Qwen2VLForConditionalGeneration.from_pretrained(
    model_path,
    torch_dtype=torch.bfloat16,
    attn_implementation="eager", # replace with "flash_attention_2" if your GPU is Ampere architecture
    device_map="auto"
)

tokenizer =  AutoTokenizer.from_pretrained(model_path)
# processor = AutoProcessor.from_pretrained(model_path)

min_pixels = 256 * 28 * 28
max_pixels = 1280 * 28 * 28
processor = AutoProcessor.from_pretrained(
     model_path,
     min_pixels=min_pixels,
     max_pixels=max_pixels,
 )

image_path ="image.jpg"

with open(image_path, "rb") as f:
    encoded_image = base64.b64encode(f.read())
decoded_image_text = encoded_image.decode('utf-8')
base64_data = f"data:image;base64,{decoded_image_text}"

messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "image",
                "image": base64_data,
            },
            {
                "type": "text",
                "text": "Trích xuất thông tin nội dung từ hình ảnh được cung cấp."
            },
        ],
    }
]

# Prepare prompt
tokenized_text = processor.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True
)

image_inputs, video_inputs = process_vision_info(messages)

inputs = processor(
    text=[ tokenized_text],
    images=image_inputs,
    videos=video_inputs,
    padding=True,
    return_tensors="pt",
)
inputs = inputs.to("cuda")

# Generation configs
generation_config =  model.generation_config
generation_config.do_sample   = True
generation_config.temperature = 1.0
generation_config.top_k       = 1
generation_config.top_p       = 0.9
generation_config.min_p       = 0.1
generation_config.best_of     = 5
generation_config.max_new_tokens     = 2048
generation_config.repetition_penalty = 1.06

# Inference
generated_ids = model.generate(**inputs, generation_config=generation_config)
generated_ids_trimmed = [
    out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
]
output_text = processor.batch_decode(
    generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
)

print(output_text[0])

References 📑

[1] Qwen team. Qwen2-VL. 2024.

[2] Bai, Jinze, et al. "Qwen-VL: A Versatile Vision-Language Model for Understanding, Localization, Text Reading, and Beyond." arXiv preprint arXiv:2308.12966 (2023).

[4] Yang, An, et al. "Qwen2 technical report." arXiv preprint arXiv:2407.10671 (2024).

[5] Chen, Zhe, et al. "Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2024.

[6] Chen, Zhe, et al. "How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites." arXiv preprint arXiv:2404.16821 (2024).

[7] Tran, Chi, and Huong Le Thanh. "LaVy: Vietnamese Multimodal Large Language Model." arXiv preprint arXiv:2404.07922 (2024).

Contact 🤝

• For correspondence regarding this work or inquiry for API trial, please contact Nguyễn Anh Nguyên at [email protected].
• Follow us on EraX Github