Model Training & Fine-Tuning AI Tools
Open-source tools for fine-tuning LLMs, diffusion models, and other AI models using LoRA, QLoRA, and full training methods.
Adapting a released checkpoint to a specific domain or behavior is the work here, and the people doing it are usually engineers holding a dataset a general purpose model handles badly. Practice splits into four camps. Adapter training built on LoRA and QLoRA occupies the practical middle, with LLaMA-Factory, ms-swift, and XTuner wrapping the same recipes in config files so one consumer GPU can carry a mid sized run. Below that sits the from scratch tier, where torchtitan, MaxText, and GPT-NeoX own the whole loop and assume a cluster with fast interconnect. A third camp handles preference and RL post training through TRL, OpenRLHF, and verl. A fourth skips gradient descent: mergekit blends existing checkpoints, while GPTQ, AWQ, and bitsandbytes trade precision for memory. What separates them is how much of the loop the operator controls versus what hardware that control costs. A newcomer should start with nanoGPT, small enough to read end to end and see what a training loop does, then move to LLaMA-Factory for real LoRA and QLoRA runs, or Kohya-ss if the target is a diffusion model, whose conventions do not carry over from LLM work. The trap to check early is licensing, which lives in the base weights rather than the trainer. Most of these repositories are Apache 2.0 or MIT, but an adapter is a derivative of the checkpoint it trained against, so a LoRA over a community licensed or research only model inherits those redistribution and commercial use limits. Check the license before committing compute.