Akshath Tiwari
← All projects

Foundational / Educational

Building a GPT-Style LLM From Scratch

Implemented a GPT-style language model end-to-end from first principles — tokenization, BPE, embeddings, attention — following Sebastian Raschka's methodology, to build real intuition beneath the frameworks.

PyTorchtiktoken

Regex tokenizer → custom vocabulary with special tokens → byte-pair encoding via tiktoken (GPT-2 vocab) → sliding-window input/target pairs → a PyTorch Dataset/DataLoader → token and positional embeddings (vocab 50,257 × dim 256). Built as a deliberate foundations exercise: the same primitives underpin every fine-tuning and inference project in this portfolio.