zhele

Zhele provides a comprehensive environment for neural network development, offering an integrated language, compiler, and IDE. Key capabilities include: • Visual block-based net construction alongside text editing • One-click GPU-based training on custom datasets • Deep introspection into net behavior using specialized "lenses" • Ready-to-use templates for modern network architectures (Llama 3, Qwen3, DeepSeek V3) • Extensive import/export options for popular frameworks (Hugging Face, PyTorch, ONNX, GGUF) This platform simplifies the complex process of building and understanding sophisticated computational models. Users can visually assemble networks from pre-defined blocks, or write code directly, with changes reflected in real-time across both views. The integrated training system allows for rapid iteration and performance monitoring on local GPUs or CPUs, providing live loss curves and detailed insights into model behavior. The environment includes a 15-lesson course designed to guide users through the process of building their own language models from scratch, explaining the "why" behind each step. Advanced debugging tools, such as a wire debugger and kernel profiler, enable users to dissect internal operations, visualize weights, computation, and memory usage. It supports seamless integration with existing ecosystems through versatile import capabilities from Hugging Face, PyTorch, and ONNX, and export options to Hugging Face, GGUF, and ONNX. Zhele is ideal for researchers, developers, and students focused on neural network construction and analysis, providing a powerful, transparent, and user-friendly platform for advanced computational modeling.