Rapidscall

Rapidscall empowers users to develop custom language models from the ground up, entirely on their Mac. Key features include: * Local, on-device model training * No coding or cloud services required * Custom tokenizer creation * Real-time training monitoring * Data privacy: no data leaves your device This robust application is designed for anyone needing to build specialized language models without external dependencies. Users can import diverse text datasets, construct a custom tokenizer using byte-level BPE, character BPE, or SentencePiece, and then configure a transformer model with options for next-word prediction or conversational formats. The platform supports optimizers like AdamW, Lion, and SGD, with adjustable learning rate schedules. Rapidscall provides comprehensive tools for monitoring the training process in real time, displaying live loss charts, gradient norms, and learning rates. Training can be paused, resumed, or stopped at any point, with progress automatically saved. The integrated early stopping mechanism helps prevent overfitting, and trained models can be immediately tested with the built-in inference feature. Built specifically for Apple Silicon, the app utilizes Metal for GPU-accelerated performance, adapting to your hardware and making your entire RAM available for training. Ideal for researchers, data scientists, and developers who prioritize data sovereignty and performance, Rapidscall ensures that all training data, tokenizers, and models remain secure and local. The free tier offers a complete exploration of the workflow, allowing training of single-layer models, while the Pro version unlocks unlimited layers for more advanced, production-quality model development.