Kalorda

An integrated fine-tuning platform for lightweight vlmOCR models

English  中文

🔥 News: Kalorda now supports fine-tuning for Deepseek-OCR-2, and specifically supports higher vLLM versions (v0.13/0.14/0.15) for running inference with Deepseek-OCR-2 (first release online).

Overview

Kalorda is a lightweight VLM OCR fine-tuning platform. The frontend is built with TypeScript + Vue3 + Vite, and the backend is built with Python + FastAPI + ms-swift + vLLM. It provides a one-stop solution for data relabeling, fine-tuning, and evaluation for mainstream lightweight VLM OCR models.

VLM OCR models are evolving rapidly. Different models have their own strengths and limitations, so real-world applications often need secondary fine-tuning to improve recognition performance in specific business scenarios. Although there are many open-source components available for data labeling, fine-tuning, and inference, there is still a lack of an integrated tool that links the entire workflow together. This makes fine-tuning work (even if it is just tool orchestration) inconvenient and challenging for non-experts. Kalorda wraps mainstream tools like ms-swift + vLLM and deeply integrates mainstream OCR models, providing an intuitive web UI that lowers the barrier to VLM OCR fine-tuning and makes operations simpler and more convenient.

Currently supported VLM OCR models:

Model Name	Model Size	Release Date	Publisher
GOT-OCR2.0	0.6B	2025	StepFun
dotsOCR	3B	2025	Xiaohongshu
Dolphin_v2	3B	Jan 2025	ByteDance
Deepseek_OCR	3B	Jan 2025	DeepSeek
PaddleOCR_VL	0.9B	Jan 2025	Baidu
HunyuanOCR	1B	Feb 2025	Tencent
Deepseek_OCR2	3B	2026	DeepSeek

More models will be integrated. PRs and issues are welcome.

Installation

Quick Install

Kalorda packages are published on PyPI. You can install directly with pip without cloning the git source code.

1. Create a virtual environment

# Create a virtual environment using conda
conda create -n kalorda python=3.12 -y

# Activate the virtual environment
conda activate kalorda

2. Install

pip install kalorda

# Or install with an Aliyun mirror
pip install kalorda -i https://mirrors.aliyun.com/pypi/simple/

3. Start

kalorda --port 8800

Optional startup parameters:

• --host: specify host address, default is 0.0.0.0
• --port: specify port, default is 8800
• --gpu-devices: specify GPU device indices (starting from 0). Default is empty, meaning all GPUs are allowed. Multiple GPUs are separated by commas, e.g. --gpu-devices 0,1,2
• --workers: specify worker process count (at least 2). Default is 2
• --log-level: specify log level, default is info

4. Login

Default admin account: admin
Default password: admin123

System and hardware requirements

• Linux OS (on Windows, please install WSL2 Ubuntu)
• Python virtual environment manager (Miniconda3 or uv recommended)
• At least one Nvidia GPU, 6GB VRAM or above; GPU driver and CUDA installed (non-Nvidia GPUs are not supported currently)
• Disk space: ?0GB or more

Source Installation

If you want to install or debug the project with frontend and backend separated, follow the steps below:

1. Clone source

git clone https://github.com/vlmOCR/Kalorda.git

2. Install and run

This project contains two parts, located under the project root: frontend/ and backend/.

Kalorda
├── backend/     # backend project
├── frontend/    # frontend project
├── LICENSE      # project license (Apache-2.0)
└── README.md    # github homepage

Install and run the backend (vLLM does not support pure Windows; the backend must run on Linux or Windows/WSL2):

# Enter the backend directory (adjust path to your actual environment)
cd /mnt/d/test/Kalorda/backend/

# Create a virtual environment using conda
conda create -n kalorda python=3.12 -y

# Activate the virtual environment
conda activate kalorda

# Install dependencies
pip install -e .[dev]

# Start (enter src/kalorda directory)
cd /mnt/d/test/Kalorda/backend/src/kalorda/
python -m main --port 8800

Install and run the frontend (requires Node.js, OS不限):

# Enter the frontend directory (adjust path to your actual environment)
cd d:/test/Kalorda/frontend/

# Install dependencies
npm install

# Open the .env.dev file in the frontend directory and set VITE_API_SERVER_URL
# to the running kalorda backend URL.
# Example: VITE_API_SERVER_URL=http://172.18.35.246:8800
# Note: update the IP address to match your backend address.

# Start
npm run dev

# Open the frontend page (default port is 8060; you can change server.port in vite.config.ts)
# Open your browser and visit http://localhost:8060

Build

Build the frontend first:

# Enter the frontend directory (adjust path to your actual environment)
cd d:/test/Kalorda/frontend/

# Run frontend build
npm run build

Built static assets will be saved under backend/src/kalorda/web_dist by default, so the backend build can include them.

Then build the backend:

# Enter the backend directory (adjust path to your actual environment)
cd /mnt/d/test/Kalorda/backend/

# Install build tool
pip install build

# Run build
python -m build

Built wheel files are saved under backend/dist by default. Example install command:

pip install kalorda-0.1.6-py3-none-any.whl

Contact

Email: postmaster@vlmocr.com

GitHub Issues: https://github.com/vlmOCR/Kalorda/issues

WeChat: lery2021

(Scan to add WeChat, note: kalorda, and you will be added to the group.)

License

Kalorda is open-sourced under the Apache-2.0 license. You are free to use, modify, and distribute this project as long as you comply with the license.

Apache-2.0

Copyright (c) 2025-present, Kalorda