BROOK - A Configurable Neural Language Model in C originally created by Sean Lane Fuller

Named for its flowing, natural way of connecting words and concepts, like a brook that connects different parts of the landscape with gentle, continuous flow.

This is a feedforward neural network that learns to predict the next word in a sequence. It uses word embeddings, multiple hidden layers with ReLU activation, dropout regularization, and various optimization techniques.

Architecture Overview: Input → Word Embeddings → Hidden Layer 1 → ... → Hidden Layer N → Output

Key Features:

• Configurable number of hidden layers
• Dynamic memory allocation for different architectures
• Xavier weight initialization for stable training
• Gradient clipping to prevent exploding gradients
• Top-k sampling for diverse text generation
• Interactive training and generation interface

Commands:

train - train with default num epochs

train N - train with N epochs

save - save current model

vocab - list all vocabulary words

tokens - list some tokens

quit - exit

or type seed words for text generation

Notes:

The current model was trained on half.txt which is half a novel that I wrote and then it was tuned on story.txt which contains the entire novel.