align-anything

Why we included this project

Fine-tuning a multimodal model so it actually follows human instructions usually means committing to one alignment method and building the data pipeline around it. Align Anything is a modular training stack that supports the main algorithms, SFT, DPO, PPO, GRPO, and rule-based RL, across image, audio, video, and any-to-any models, so you can compare approaches without rebuilding your workflow. It ships working scripts for Qwen3, DeepSeek, Janus, and MiniCPM-o, plus tutorials that show how to take a text-only model and add image input or set up DPO training. The scaffolding for NVIDIA GPUs, Huawei Ascend NPUs, and Slurm clusters is already there, which makes it a practical place to start real alignment experiments rather than a toy.