EvoNet-Studio Pro is a cutting-edge, web-based (Gradio) interface designed to make Supervised Fine-Tuning (SFT), Reinforcement Learning (RLHF), and Model Deployment accessible, fast, and secure. Built on top of the robust vietnamese-ai framework and optimized with Unsloth, this studio allows you to build, align, test, and deploy state-of-the-art Large Language Models (LLMs) and Vision-Language Models (VLMs) right from your local machine.
Move beyond basic SFT. EvoNet-Studio supports the holy trinity of AI alignment:
- DPO (Direct Preference Optimization)
- ORPO (Odds Ratio Preference Optimization) - Highly VRAM efficient!
- KTO (Kahneman-Tversky Optimization)
- Agent Tuning: Train models to use external APIs and Tools via standard ReAct or function-calling formats (ChatML, Llama-3-Instruct).
- Vision-Language Models (VLM): Built-in toggle to fine-tune multi-modal models like LLaVA or Qwen-VL using
FastVisionModel.
Stop scraping data! Use the integrated Google Gemini API (or OpenAI/Anthropic/Local LLMs) to dynamically generate high-quality .jsonl datasets (Alpaca/ChatML format) based on your custom prompt.
Test your fine-tuned model's ability to extract facts from private documents:
- Upload
.pdfor.txtfiles. - The built-in Local Vector Database (
sentence-transformers) chunks and embeds the document. - Chat with your model securely offline and verify its RAG capabilities.
Ditch traditional Perplexity metrics. Use our built-in LLM Arena to blindly test two models side-by-side. Your votes are automatically saved as a high-quality DPO dataset (outputs/arena_dpo_dataset.jsonl). You can also use Gemini, GPT-4, or Claude to automatically grade your model's responses.
Stop struggling with prompt engineering! Enter a basic idea (e.g., "A customer service bot") and our built-in Auto-Prompt Optimizer will use advanced Meta-Prompting to generate a highly structured, professional System Prompt (Role, Objective, Rules, Format) instantly.
Garbage in, garbage out! Use the built-in Data Cleaner to instantly analyze your raw .jsonl datasets. It automatically detects and removes duplicate records, identifies missing keys, strips out overly short responses, and estimates total tokens to ensure your dataset is perfectly healthy before wasting VRAM on training.
Stop guessing if your model is smart! We integrated lm-eval (HuggingFace Evaluation Harness) directly into EvoNet-Studio. Select your fine-tuned model, choose tests like MMLU, GSM8k, or HellaSwag, and run a rigorous academic benchmark locally to get objective scores.
The Holy Grail of AI: A closed feedback loop! When deploying the model via our Native API, all chats are logged. External apps can send Thumbs Up/Down feedback to the /v1/feedback endpoint. Inside the Telemetry Studio, you can view bad responses, correct them manually, and instantly push them to your DPO Dataset to make your next model version smarter!
Once trained, deploy your model as a production-grade API server (http://localhost:8000/v1/chat/completions) with a single click.
- vLLM Engine: For those with powerful GPUs, deploy with PagedAttention for maximum throughput.
- FastAPI Engine: Native CPU/GPU fallback using standard Transformers (supports Production Telemetry).
Never suffer from Catastrophic Forgetting again! Instead of overwriting your base model, train isolated LoRA experts (e.g., Code, Math, Law). In the API Deployment tab, load your Base Model and pass a JSON dictionary of your LoRA adapters. The Native Server acts as a Mixture-of-Experts (MoE) Router, dynamically analyzing incoming prompts and hot-swapping the correct LoRA adapter in 0.01s before generating a response. Serve 50 experts from a single GPU without OOM! You can easily manage all your trained adapters using the dedicated LoRA Manager Tab.
EvoNet-Studio is future-proof. Click the "Fetch Latest Models" button to instantly query the HuggingFace API and pull down the top trending models (e.g., latest Llama, Qwen, Gemma variants from Unsloth) straight into your training dropdowns.
Merge your trained LoRA adapters directly into the Base Model and instantly push the standalone model to the HuggingFace Hub to share with the world. You can also export the model to GGUF format and push it directly to Ollama with a single click to run on local CPU/Edge devices!
Train your model on massive amounts of raw text (e.g., medical textbooks, legal documents) without formatting them into Q&A. The new CPT mode allows you to inject deep domain knowledge natively into the base model before you fine-tune it for conversation.
Never lose your precious datasets or fine-tuned weights! The Cloud Sync tab allows you to compress your entire outputs/ folder and securely back it up to a Private HuggingFace Dataset repository. Switch computers? Just click "Restore Workspace" to pull everything down and pick up exactly where you left off.
For corporate environments and heavy workloads, the EvoNet-Studio Enterprise Edition includes:
- API Key Gateway & Rate Limiting: Prevent D-DoS and control token usage with internal API keys.
- SSO & OAuth2 Integration: 1-Click login via Google/Microsoft Azure AD for entire teams.
- Multi-Node Distributed Training: Seamlessly enable DeepSpeed ZeRO-3 to distribute heavy training across multiple physical GPU nodes.
- GPU Cost Dashboard: Real-time tracking of GPU uptime and estimated compute costs (in USD).
- Canary & A/B Deployments: Safely route a percentage of API traffic to a Challenger Model to test in production.
- Hardware Lock & Auto PII Masking: Compliance and security for sensitive data.
We provide an automated setup script that creates an optimized virtual environment:
# 1. Clone the repository
git clone https://github.com/phonghhd/EvoNet-Studio.git
cd EvoNet-Studio
# 2. Run the setup script
bash setup.sh
# 3. Activate the environment
source venv/bin/activate
# 4. Start the Studio!
Bạn có thể khởi động EvoNet-Studio Pro bằng nhiều cách tùy theo nhu cầu bảo mật:
**Chạy Cục bộ (Local - Khuyên dùng cho cá nhân):**
```bash
python3 app.py(Chỉ truy cập được từ máy tính của bạn thông qua http://localhost:7860)
Chạy trên Cloud/Kaggle (Cần Public Link):
python3 app.py --share(Tạo đường link .gradio.live. Lưu ý: Ai có link cũng có thể vào được!)
Chạy trên Cloud với Mật khẩu Bảo mật (Khuyên dùng):
python3 app.py --share --auth admin:matkhau123(Tạo link Public nhưng bị khóa bởi màn hình Đăng nhập. Thay đổi admin:matkhau123 thành tài khoản và mật khẩu của bạn).
For enterprise users looking to deploy EvoNet-Studio on cloud instances (AWS EC2, GCP, Azure) or local servers, we provide a production-ready Docker setup.
# 1. Clone the repository
git clone https://github.com/phonghhd/EvoNet-Studio.git
cd EvoNet-Studio
# 2. Deploy with Docker Compose (requires NVIDIA Container Toolkit)
docker-compose up -d --buildWhat this does:
- Automatically builds an isolated container using
pytorch:2.2.1-cuda12.1as the base image. - Passes GPU passthrough to the container (
capabilities: [gpu]). - Exposes port
7860for the Studio UI and port8000for the API Deployment server. - Mounts local
/outputsand/datasetdirectories to prevent data loss when the container restarts.
EvoNet-Studio intelligently detects your hardware capabilities:
- GPU Present: The system will dynamically load
Unslothkernels to perform 4-bit quantization fine-tuning, resulting in 2x faster training and 70% less VRAM usage. - CPU Only: The system gracefully falls back to native HuggingFace
transformersandtrl, allowing you to develop, test, and run the entire pipeline even without a dedicated graphics card.
Tài liệu hướng dẫn chi tiết dành cho các phiên bản của EvoNet-Studio:
EvoNet-Studio is constantly evolving. In the upcoming Q3/Q4 releases, we are planning to introduce:
- Apple Silicon (MPS) Optimization: Enhanced hardware acceleration for MacOS users without NVIDIA cards.
- Llama.cpp WebUI Integration: Connect directly to local model management interfaces for smoother execution.
- Multi-Tenant Architecture (Enterprise): True SaaS-ready architecture allowing a single installation to serve multiple isolated departments.
- Advanced RAG Evaluation Suites (Enterprise): Built-in RAG measurement tools (RAGAS, TruLens) directly on the dashboard to score AI factual accuracy.
Contributions are welcome! If you'd like to add support for new Alignment algorithms, better UI elements, or new Deployment engines, feel free to submit a Pull Request.
This project is licensed under the MIT License - see the LICENSE file for details.
Nếu bạn thấy EvoNet-Studio Pro hữu ích cho công việc, học tập của bạn, hãy cân nhắc ủng hộ (donate) để tiếp thêm động lực cho chúng tôi duy trì và phát triển thêm nhiều siêu tính năng mới trong tương lai!
Mọi sự đóng góp dù nhỏ nhất đều là nguồn động viên vô giá đối với những người làm mã nguồn mở.
- Ngân hàng: MB BANK
- Số tài khoản: 0948622661
- Tên người thụ hưởng: HUYNH DUONG PHONG
- Nội dung: Donate EvoNet Studio
Cảm ơn bạn rất nhiều vì đã đồng hành cùng sự phát triển của Trí tuệ Nhân tạo Việt Nam! 🇻🇳