Align-Anything
libraryPKU-Alignment's all-modality alignment framework (4.7K★; Yaodong Yang, Institute for AI): trains any-to-any models across text, image, audio, and video with RLHF/DPO-family algorithms behind unified interfaces — the backbone of the group's later releases and companion to the standard PKU-SafeRLHF safety-preference dataset (~15K monthly HF downloads) and the InterMT multi-turn interleaved preference data (NeurIPS 2025 Spotlight).