Show HN: Neurolume – Fine-tune LLMs on a Mac and watch what changes per layer (github.com)

Neurolume, a new tool launched in early alpha, enables users to fine-tune large language models (LLMs) on Mac devices with a simple interface while providing real-time insights into model changes at each layer. Unlike traditional fine-tuning applications that report only final metrics like "loss = 1.23," Neurolume offers a dual-functionality where users can both train the model and visually monitor internal modifications during the process. This capability emphasizes interpretability, allowing developers to observe how specific layers evolve and contribute to outputs, facilitating a deeper understanding of model behavior. The tool features guided LoRA training, which restricts modifications to the last 16 layers to avoid significant disruption, while offering validation loss assessments and comparative analyses with original models. Neurolume's real-time visualizations include a "live brain view," 3D layer mappings, and a logit lens that clarifies each layer's predictive contributions during token generation. This provides valuable insights for ML practitioners seeking to customize model behavior effectively and encourages ongoing exploration of insights into AI behaviors, ultimately enhancing the training process and fostering better model reliability and adaptability.