Language Model Builder Go84jh

Language Model Builder is a free Mac app that combines an approachable introduction to language models with a local training workbench. Its interactive textbook explains next-token prediction, tokenization, embeddings, attention, transformers, training data, loss, gradient descent, supervised fine-tuning, and direct preference optimization. Interactive examples let learners build a tokenizer, explore embedding space, inspect attention, and watch a small model learn. The native workbench runs the full model-building process on an Apple Silicon Mac using hardware-accelerated MLX, without requiring additional tools. Users choose pre-training and fine-tuning data, run pre-training, sample the base model, apply supervised fine-tuning or DPO, and chat with the resulting model. During a run, the app displays loss, throughput, validation, saved samples, and checkpoint history. Training can be paused and resumed from saved checkpoints. The final chat interface includes inspection features such as token probabilities and unselected tokens. Models, datasets, and training history remain on the Mac, and generated models can be shared as safetensors files. The app is aimed at curious beginners, educators, and technically minded learners using Apple Silicon Macs with macOS 15 or later.