V0.5.0 of PSSA: a non-transformer language model (github.com)

The first tagged release of PSSA version 0.5.0 has been announced, marking a significant milestone in the development of this non-transformer language model. This update includes a comprehensive training core designed to optimize performance through innovative features like parameter-matched transformer baselines and resumable learning-rate horizons. The introduction of GPU backends utilizing native CUDA/cuBLAS and WebGPU enhances computational efficiency, while advanced techniques such as gradient clipping and episodic memory value caps improve reliability during training. This release is notable for the AI/ML community as it demonstrates advancements in model distillation with the Qwen3-to-PSSA converter, employing a rigorously validated loss function to ensure optimal learning. The revamped terminal user interface (TUI) and updated documentation enhance user accessibility and project navigation, while also facilitating contributions from the community. Overall, version 0.5.0 reflects a concerted effort to bolster reliability, ease of use, and technical prowess within the framework, paving the way for more sophisticated applications in the AI landscape.