Nano LM Studio (github.com)

NanoLM Studio has launched as a comprehensive local desktop tool designed to enable users to build small decoder-only language models from customizable corpuses. This release combines several essential functions including document ingestion, cleaning, ByteLevel BPE tokenization, and model training within a single Tk application, enhancing the workflow with a five-tab interface that aims to resolve issues seen in previous iterations. Users can train models with various parameter settings, manage document libraries, and utilize advanced features like attention visualization and real-time metrics. This significant update addresses reliability concerns and introduces a more robust training framework, ensuring accurate model parameter tracking and validation. It supports flexible configurations, with six customizable presets ranging from Tiny to XXL, accommodating a spectrum of hardware capabilities. The project emphasizes user control over data and model settings, including the management of document types and isolated directory paths for safety. As an open-source tool under the MIT license, NanoLM Studio not only serves as an educational platform but also lays a solid foundation for future enhancements, such as mult-project workspaces and advanced inference capabilities, making it a pivotal addition to the AI/ML community portfolio.