/ ML, DL & GenAI / Phase 7 Phase 7 Days 129-148
Transformers and LLM Internals Self-attention, transformer blocks, BERT, GPT, T5, tokenizers, scaling laws, pretraining, decoding, HF internals, PEFT, tiny GPT
Phase Goal Build the transformer stack from the inside out: attention, tokenizers, GPT-style models, scaling, generation, fine-tuning, and Hugging Face parity.
Day 129: Transformer Architecture Map
Evaluation & Failure Modes Architecture understanding: the five mistakes beginners make when reading a transformer paper. Inspect individual errors, slices, prompts, or traces — never trust a single average score Catalogue leakage, overfitting, bias, hallucination, latency, cost, or operational risks for this technique Write the next experiment from the failure mode you observed Close the notebook and reconstruct one mechanism; log help used and schedule a delayed changed-task check, initially around 1/3/7/14/30 days Keep ordinary review within 20 minutes; if weak prerequisites accumulate, pause new material and repair one dependency Record two competing explanations and the smallest experiment that distinguishes them; distinguish a negative result from an implementation failure
Day 130: QKV Self-Attention From Scratch
Evaluation & Failure Modes Attention shape bugs: the four shape mistakes that produce silent garbage. Inspect individual errors, slices, prompts, or traces — never trust a single average score Catalogue leakage, overfitting, bias, hallucination, latency, cost, or operational risks for this technique Write the next experiment from the failure mode you observed Close the notebook and reconstruct one mechanism; log help used and schedule a delayed changed-task check, initially around 1/3/7/14/30 days Keep ordinary review within 20 minutes; if weak prerequisites accumulate, pause new material and repair one dependency Record two competing explanations and the smallest experiment that distinguishes them; distinguish a negative result from an implementation failure
Day 131: Multi-Head Attention
Evaluation & Failure Modes Head behaviour: heads that specialise, heads that die, and pruning intuition. Inspect individual errors, slices, prompts, or traces — never trust a single average score Catalogue leakage, overfitting, bias, hallucination, latency, cost, or operational risks for this technique Write the next experiment from the failure mode you observed Close the notebook and reconstruct one mechanism; log help used and schedule a delayed changed-task check, initially around 1/3/7/14/30 days Keep ordinary review within 20 minutes; if weak prerequisites accumulate, pause new material and repair one dependency Record two competing explanations and the smallest experiment that distinguishes them; distinguish a negative result from an implementation failure
Day 132: Positional Encoding and RoPE
Evaluation & Failure Modes Position failures at long context and the patches the field has shipped. Inspect individual errors, slices, prompts, or traces — never trust a single average score Catalogue leakage, overfitting, bias, hallucination, latency, cost, or operational risks for this technique Write the next experiment from the failure mode you observed Close the notebook and reconstruct one mechanism; log help used and schedule a delayed changed-task check, initially around 1/3/7/14/30 days Keep ordinary review within 20 minutes; if weak prerequisites accumulate, pause new material and repair one dependency Record two competing explanations and the smallest experiment that distinguishes them; distinguish a negative result from an implementation failure
Day 133: Transformer Block From Scratch
Evaluation & Failure Modes Block stability across depth: when gradients explode and the norm placement that fixes it. Inspect individual errors, slices, prompts, or traces — never trust a single average score Catalogue leakage, overfitting, bias, hallucination, latency, cost, or operational risks for this technique Write the next experiment from the failure mode you observed Close the notebook and reconstruct one mechanism; log help used and schedule a delayed changed-task check, initially around 1/3/7/14/30 days Keep ordinary review within 20 minutes; if weak prerequisites accumulate, pause new material and repair one dependency Record two competing explanations and the smallest experiment that distinguishes them; distinguish a negative result from an implementation failure
Day 134: Encoder Models and BERT
Evaluation & Failure Modes Encoder limits. Where BERT-family models still beat decoders in 2026. Inspect individual errors, slices, prompts, or traces — never trust a single average score Catalogue leakage, overfitting, bias, hallucination, latency, cost, or operational risks for this technique Write the next experiment from the failure mode you observed Close the notebook and reconstruct one mechanism; log help used and schedule a delayed changed-task check, initially around 1/3/7/14/30 days Keep ordinary review within 20 minutes; if weak prerequisites accumulate, pause new material and repair one dependency Record two competing explanations and the smallest experiment that distinguishes them; distinguish a negative result from an implementation failure
Day 135: Decoder-Only GPT From Scratch
Evaluation & Failure Modes Generation behaviour: the four sampling pathologies (loops, drift, hallucination, repetition). Inspect individual errors, slices, prompts, or traces — never trust a single average score Catalogue leakage, overfitting, bias, hallucination, latency, cost, or operational risks for this technique Write the next experiment from the failure mode you observed Close the notebook and reconstruct one mechanism; log help used and schedule a delayed changed-task check, initially around 1/3/7/14/30 days Keep ordinary review within 20 minutes; if weak prerequisites accumulate, pause new material and repair one dependency Record two competing explanations and the smallest experiment that distinguishes them; distinguish a negative result from an implementation failure
Day 136: Encoder-Decoder Models and T5
Evaluation & Failure Modes Conditional generation errors: source copying, hallucination, and prefix sensitivity. Inspect individual errors, slices, prompts, or traces — never trust a single average score Catalogue leakage, overfitting, bias, hallucination, latency, cost, or operational risks for this technique Write the next experiment from the failure mode you observed Close the notebook and reconstruct one mechanism; log help used and schedule a delayed changed-task check, initially around 1/3/7/14/30 days Keep ordinary review within 20 minutes; if weak prerequisites accumulate, pause new material and repair one dependency Record two competing explanations and the smallest experiment that distinguishes them; distinguish a negative result from an implementation failure
Day 137: BPE and WordPiece Tokenizers From Scratch
Evaluation & Failure Modes Tokenization surprises: numbers, code, emojis, and non-Latin scripts. Inspect individual errors, slices, prompts, or traces — never trust a single average score Catalogue leakage, overfitting, bias, hallucination, latency, cost, or operational risks for this technique Write the next experiment from the failure mode you observed Close the notebook and reconstruct one mechanism; log help used and schedule a delayed changed-task check, initially around 1/3/7/14/30 days Keep ordinary review within 20 minutes; if weak prerequisites accumulate, pause new material and repair one dependency Record two competing explanations and the smallest experiment that distinguishes them; distinguish a negative result from an implementation failure
Day 138: Scaling Laws and Pretraining Data
Evaluation & Failure Modes Data contamination: how it inflates benchmarks and how to detect it. Inspect individual errors, slices, prompts, or traces — never trust a single average score Catalogue leakage, overfitting, bias, hallucination, latency, cost, or operational risks for this technique Write the next experiment from the failure mode you observed Close the notebook and reconstruct one mechanism; log help used and schedule a delayed changed-task check, initially around 1/3/7/14/30 days Keep ordinary review within 20 minutes; if weak prerequisites accumulate, pause new material and repair one dependency Record two competing explanations and the smallest experiment that distinguishes them; distinguish a negative result from an implementation failure
Day 139: Training Large Language Models
Evaluation & Failure Modes Training bottlenecks: communication, memory, and the three things that kill throughput. Inspect individual errors, slices, prompts, or traces — never trust a single average score Catalogue leakage, overfitting, bias, hallucination, latency, cost, or operational risks for this technique Write the next experiment from the failure mode you observed Close the notebook and reconstruct one mechanism; log help used and schedule a delayed changed-task check, initially around 1/3/7/14/30 days Keep ordinary review within 20 minutes; if weak prerequisites accumulate, pause new material and repair one dependency Record two competing explanations and the smallest experiment that distinguishes them; distinguish a negative result from an implementation failure
Day 140: Efficient Attention and Long Context
Evaluation & Failure Modes Long-context failure modes: lost-in-the-middle and the prompt patterns that trigger it. Inspect individual errors, slices, prompts, or traces — never trust a single average score Catalogue leakage, overfitting, bias, hallucination, latency, cost, or operational risks for this technique Write the next experiment from the failure mode you observed Close the notebook and reconstruct one mechanism; log help used and schedule a delayed changed-task check, initially around 1/3/7/14/30 days Keep ordinary review within 20 minutes; if weak prerequisites accumulate, pause new material and repair one dependency Record two competing explanations and the smallest experiment that distinguishes them; distinguish a negative result from an implementation failure
Day 141: Generation Decoding Algorithms
Evaluation & Failure Modes Bad generations: the symptoms that point to each decoding parameter. Inspect individual errors, slices, prompts, or traces — never trust a single average score Catalogue leakage, overfitting, bias, hallucination, latency, cost, or operational risks for this technique Write the next experiment from the failure mode you observed Close the notebook and reconstruct one mechanism; log help used and schedule a delayed changed-task check, initially around 1/3/7/14/30 days Keep ordinary review within 20 minutes; if weak prerequisites accumulate, pause new material and repair one dependency Record two competing explanations and the smallest experiment that distinguishes them; distinguish a negative result from an implementation failure
Day 142: Perplexity, Benchmarks, Contamination
Evaluation & Failure Modes Benchmark trust. The three questions to ask before quoting any leaderboard number. Inspect individual errors, slices, prompts, or traces — never trust a single average score Catalogue leakage, overfitting, bias, hallucination, latency, cost, or operational risks for this technique Write the next experiment from the failure mode you observed Close the notebook and reconstruct one mechanism; log help used and schedule a delayed changed-task check, initially around 1/3/7/14/30 days Keep ordinary review within 20 minutes; if weak prerequisites accumulate, pause new material and repair one dependency Record two competing explanations and the smallest experiment that distinguishes them; distinguish a negative result from an implementation failure
Day 143: Prompting and In-Context Learning Internals
Evaluation & Failure Modes Prompt brittleness and the patterns that make a prompt portable across model versions. Inspect individual errors, slices, prompts, or traces — never trust a single average score Catalogue leakage, overfitting, bias, hallucination, latency, cost, or operational risks for this technique Write the next experiment from the failure mode you observed Close the notebook and reconstruct one mechanism; log help used and schedule a delayed changed-task check, initially around 1/3/7/14/30 days Keep ordinary review within 20 minutes; if weak prerequisites accumulate, pause new material and repair one dependency Record two competing explanations and the smallest experiment that distinguishes them; distinguish a negative result from an implementation failure
Day 144: Hugging Face Transformers Internals
Evaluation & Failure Modes Library pitfalls: silent default changes between versions, and how to pin them. Inspect individual errors, slices, prompts, or traces — never trust a single average score Catalogue leakage, overfitting, bias, hallucination, latency, cost, or operational risks for this technique Write the next experiment from the failure mode you observed Close the notebook and reconstruct one mechanism; log help used and schedule a delayed changed-task check, initially around 1/3/7/14/30 days Keep ordinary review within 20 minutes; if weak prerequisites accumulate, pause new material and repair one dependency Record two competing explanations and the smallest experiment that distinguishes them; distinguish a negative result from an implementation failure
Day 145: Fine-Tuning Encoder and Decoder Models
Evaluation & Failure Modes Fine-tuning overfit: how to spot it from the loss curve in 30 seconds. Inspect individual errors, slices, prompts, or traces — never trust a single average score Catalogue leakage, overfitting, bias, hallucination, latency, cost, or operational risks for this technique Write the next experiment from the failure mode you observed Close the notebook and reconstruct one mechanism; log help used and schedule a delayed changed-task check, initially around 1/3/7/14/30 days Keep ordinary review within 20 minutes; if weak prerequisites accumulate, pause new material and repair one dependency Record two competing explanations and the smallest experiment that distinguishes them; distinguish a negative result from an implementation failure
Day 146: Parameter-Efficient Fine-Tuning
Evaluation & Failure Modes Adapter quality. The signs your LoRA rank is too low or too high. Inspect individual errors, slices, prompts, or traces — never trust a single average score Catalogue leakage, overfitting, bias, hallucination, latency, cost, or operational risks for this technique Write the next experiment from the failure mode you observed Close the notebook and reconstruct one mechanism; log help used and schedule a delayed changed-task check, initially around 1/3/7/14/30 days Keep ordinary review within 20 minutes; if weak prerequisites accumulate, pause new material and repair one dependency Record two competing explanations and the smallest experiment that distinguishes them; distinguish a negative result from an implementation failure
Day 147: Reproducing GPT-2 Style Components
Evaluation & Failure Modes Implementation parity. The numerical tests that prove your reimplementation is correct. Inspect individual errors, slices, prompts, or traces — never trust a single average score Catalogue leakage, overfitting, bias, hallucination, latency, cost, or operational risks for this technique Write the next experiment from the failure mode you observed Close the notebook and reconstruct one mechanism; log help used and schedule a delayed changed-task check, initially around 1/3/7/14/30 days Keep ordinary review within 20 minutes; if weak prerequisites accumulate, pause new material and repair one dependency Record two competing explanations and the smallest experiment that distinguishes them; distinguish a negative result from an implementation failure
Day 148: Capstone: Tiny GPT and Transformer Internals
Evaluation & Failure Modes LLM internals capstone quality. The README that turns this into a portfolio piece. Inspect individual errors, slices, prompts, or traces — never trust a single average score Catalogue leakage, overfitting, bias, hallucination, latency, cost, or operational risks for this technique Write the next experiment from the failure mode you observed Close the notebook and reconstruct one mechanism; log help used and schedule a delayed changed-task check, initially around 1/3/7/14/30 days Keep ordinary review within 20 minutes; if weak prerequisites accumulate, pause new material and repair one dependency Record two competing explanations and the smallest experiment that distinguishes them; distinguish a negative result from an implementation failure
Capstone: Tiny GPT and Transformer Internals Build a tiny GPT-style model, implement or inspect the tokenizer, train or fine-tune it on a small corpus, compare with Hugging Face, and write an evaluation report.
After this phase, you'll be able to:
Implement self-attention and transformer blocks Build or inspect BPE tokenizers and GPT-style training loops Fine-tune and evaluate encoder and decoder models Understand scaling, long context, decoding, and PEFT tradeoffs You can connect Karpathy-style GPT internals with CS224N-style transformer rigor and practical HF workflows. Gate: explain without notes, build one independent artifact, diagnose a deliberate failure, and repeat a changed task after a delay. Record help and repair missing prerequisites.