Firefly

Why we included this project

Firefly is a practical training toolkit for teams that want to fine-tune an open model rather than start from scratch. It handles pretraining, instruction fine-tuning, and DPO preference tuning through config-driven scripts, and the supported model list is wide: Qwen, Llama, Yi, Gemma, MiniCPM, DeepSeek, Mistral, Mixtral, Baichuan, ChatGLM, InternLM, and more. Because the training templates line up with each official chat model, you can swap in a different architecture without reworking the data format yourself. The project is a good fit when GPU memory is limited: QLoRA is the recommended path for tight budgets, validated on the Open LLM Leaderboard, and an Unsloth integration brings Llama3-8B training down to roughly 7.75 GB of VRAM with faster runs. That mix of model coverage and low-hardware options makes it a sensible first stop for small teams adapting a base model.