SpencerGPT
A GPT-architecture Transformer language model built entirely from scratch in PyTorch.
Architecture
- Multi-head self-attention (4 heads)
- 4 Transformer blocks with residual connections
- 64-dimensional embeddings
- Causal masking, LayerNorm, Dropout
- Character-level tokenizer
- Trained on CPU (Apple Intel Mac)
Built by
colonel1223 — built from first principles, no pre-existing framecolks used.
