The arc of this phase is the arc of modern AI: text becomes vectors (TF-IDF → word2vec), vectors learn to talk to each other (attention), and the machinery that lets them talk at scale is the Transformer — which you'll assemble conceptually, then build for real (BYO-8) and fine-tune with Hugging Face. Covers the Transformer core and the bridge into Phases 6–6B.
Fine-tune a transformer on a task that matters to you — the natural pick for your security track is a phishing-email classifier (financial-news sentiment works too) — and ship it as a Gradio demo on Hugging Face Spaces. Finish the from-scratch pair: BYO-7 (BPE tokenizer) and BYO-8 (mini-GPT) — after them, nothing in a transformer is a black box. Close by reading "Attention Is All You Need" with the 3-pass method: it will feel like reviewing, not deciphering. Ready for Phase 6 when you can fine-tune + deploy a transformer and explain attention, positional encoding and residuals end-to-end.
Full guide: phase-5-nlp-transformers.md ↗