Humigence v2 Release
Hi everyone!
I'm excited to announce the public release of Humigence v2, an open-source MLOps toolkit that makes supervised fine-tuning of LLMs fast, simple, and GPU-efficient. I am a complete n00b (I don't write code), but I wanted to get into the AI world and embark on some MLOps processes - starting with finetuning. It was quite a continuous back and forth. As a result, I thought it'd be nice if there was a step by step process to getting some MLOps processes done. I decided to start with fine-tuning.
Humigence wraps the Unsloth library with a user-friendly interactive CLI wizard, enabling both beginners and power users to fine-tune models on single GPU or multi-GPU (dual RTX 5090) setups with zero boilerplate code.
Features
• 🔧 Interactive CLI Wizard with Basic and Advanced modes (advanced is yet to be released)
• ⚡ Dual-GPU Training with torchrun + NCCL
• 🦥 Unsloth Integration for QLoRA/LoRA fine-tuning (4-bit & 16-bit)
• 📊 Training Summaries with loss curves, overfitting detection, and metrics
• 📁 Config Snapshots for full reproducibility
• 🖥️ Automatic GPU Detection (single or multi-GPU)
• ✅ LoRA adapter saving with optional merged weights
🔬 Requirements
• Python 3.10+
• CUDA 12.1+
• PyTorch 2.1+
• GPUs with at least 16GB VRAM (tested on dual RTX 5090s)
See requirements.txt for full dependency list.
🤝 Contributing
We welcome PRs, feedback, and issues! (Don't be too harsh lol)
👉 https://github.com/loladebabalola/humigencev2
🌍 Where to Find Us
• GitHub: loladebabalola/humigencev2
• Hugging Face Model Card: Humigence v2
🙏 Thanks
Thanks to the Hugging Face & Unsloth communities for providing the foundation that made this possible.
We hope Humigence v2 helps more people fine-tune LLMs efficiently on their own hardware!