Character-Level Transformer

I built a small autoregressive Transformer from scratch in PyTorch while following Andrej Karpathy's educational GPT implementation. I trained it on Shakespeare's Coriolanus and published the checkpoint and model notes on Hugging Face.

This is an educational model, not a production language model. Its value is the hands-on implementation of causal self-attention, masking, residual connections, normalization, training, and evaluation.

Architecture

  • 6 Transformer blocks
  • 6 attention heads
  • 384-dimensional embeddings
  • 256-character context
  • 4× feed-forward expansion
  • ReLU activations, LayerNorm, residual connections, and dropout

Training result

  • Training loss: 0.8848
  • Validation loss: 1.5879
  • Character-level next-token prediction

Stack

  • Python and PyTorch
  • Causal multi-head self-attention
  • Language-model training and evaluation

Link

View the model on Hugging Face