Training Explorer Launches Interactive 14-Course Explorer for LLM Training

Training Explorer released an interactive explorer that turns large language model training into a playable, step-by-step course, showing token, matrix, attention weight, residual, FFN, logits and loss changes at each step. The tool offers 14 courses covering tokenization, backpropagation, optimizers, distributed training, memory, inference and fine-tuning alignment, with derivations, numerical examples, PyTorch code, common errors and self-test questions. A sample lesson uses a hand-calculable model with 4 tokens and 2 attention heads across 12 computation steps, and progress is stored in the browser without an account.
Event version 1.
Key insights
Playable model internals
Training Explorer lets users play, single-step, rewind, jump and reset through model computation events. Each step displays internal matrices and probabilities, including token, matrix, attention weight, residual, FFN, logits and loss changes.
Hand-calculable sample lesson
A sample lesson uses a small model with 4 tokens and 2 attention heads across 12 computation steps to the output. The numbers are actually computed rather than drawn as heatmaps.
14 courses with full materials
The explorer includes 14 courses spanning tokenization, backpropagation, optimizers, distributed training, memory, inference and fine-tuning alignment. Each chapter provides derivations, numerical examples, PyTorch code, common errors, self-test questions and links to original papers and official documentation.
Browser-based progress tracking
Learning progress is stored in the browser and requires no account.