Phase 4 Cheat Sheet

Phase 4: Large Language Models — Cheat Sheet

1. LLM Foundations

Concept	Summary
LLM	Neural network trained on internet-scale text to predict next token
Autoregressive	Generates one token at a time, each depending on all previous
Tokenization	Text → subword tokens → numerical IDs (vocabulary: 50K-200K)
Context Window	Max tokens model can process (4K → 128K → 1M)
Emergent Abilities	Capabilities that appear at scale (reasoning, in-context learning)

2. Transformer Architecture

flowchart LR
    subgraph BLOCK["One Transformer Decoder Block"]
        IN["Input"] --> ATTN["Multi-Head\nSelf-Attention"]
        IN --> RESID1["+"]
        ATTN --> RESID1
        RESID1 --> LN1["LayerNorm"]
        LN1 --> FFN["Feed-Forward\nNetwork"]
        LN1 --> RESID2["+"]
        FFN --> RESID2
        RESID2 --> LN2["LayerNorm"]
        LN2 --> OUT["Output"]
    end

Component	Purpose
Self-Attention	Computes relationships between all token pairs
QKV	Query (what I need), Key (what I offer), Value (what I carry)
Multi-Head	8-96 parallel attention patterns per layer
Positional Encoding	Sinusoidal/learned — adds order information
FFN	Two-layer MLP (typically 4x hidden dim)
Residual Connections	Skip connections enabling 12-100+ layer depth
LayerNorm	Stabilizes activations, enables training

3. Training Pipeline

Stage 1: Pretraining ───→ Stage 2: SFT ───────────→ Stage 3: Alignment
(Trillions of tokens)     (100K-1M examples)        (RLHF or DPO)
     │                          │                         │
     ▼                          ▼                         ▼
 Base Model              Instruction Model          Aligned Model
(continues text)         (follows prompts)         (helpful + safe)

Stage	Data	Objective	Result
Pretraining	Raw internet text	Next token prediction	Base model
SFT	Instruction-response pairs	Language modeling on responses	Instruction model
RLHF	Human preference rankings	PPO against reward model	Aligned model
DPO	Preference pairs	Direct preference optimization	Aligned model

4. Inference & Decoding

Strategy	Behavior	Use Case
Greedy	Always pick most likely token	Factual answers, code
Beam Search	Maintain top-N sequences	Translation, summarization
Temperature	Scale logits (0=deterministic, >0=random)	Creative writing
Top-K	Sample from K most likely tokens	Balanced creativity
Top-P	Sample from cumulative probability P	Adaptive diversity

Typical Values

Temperature: 0.1-0.3 (factual), 0.7-0.9 (creative)
Top-K: 40-50
Top-P: 0.9-0.95

5. Production Considerations

Feature	How It Works
Streaming	Server-Sent Events (SSE) — tokens sent as generated
Function Calling	LLM outputs JSON with tool name + arguments
Structured Output	Grammar-based constrained decoding for valid JSON
Hallucination Mitigation	RAG, temperature control, prompting, validation

GPT Model Evolution

Model	Year	Params	Context	Key Innovation
GPT-1	2018	117M	512	Transformer decoder for language modeling
GPT-2	2019	1.5B	1024	Zero-shot generalization
GPT-3	2020	175B	2048	In-context learning, emergent abilities
GPT-3.5	2022	175B	4K/16K	Instruction tuning + RLHF
GPT-4	2023	~1.8T	8K/32K/128K	Multimodal, reasoning
GPT-4o	2024	—	128K	Real-time audio, vision, text

Quick Reference: Common LLM Families

Family	Open?	Best For
GPT (OpenAI)	No	General purpose, coding, reasoning
Claude (Anthropic)	No	Long documents, safety, analysis
Gemini (Google)	No	Very long context, multimodal
Llama (Meta)	Yes	Self-hosting, research
Mistral	Partial	Efficiency, multilingual
DeepSeek	Yes	Cost-effective, coding
Qwen (Alibaba)	Yes	Coding, math, Chinese/English