📚 Complete 20-week study plan: Learn LLMs from First Principles

README.md

@@ -0,0 +1,980 @@
+---
+title: LLM from First Principles
+emoji: 🧠
+colorFrom: blue
+colorTo: purple
+sdk: static
+pinned: true
+tags:
+  - llm
+  - tutorial
+  - learning-path
+  - transformers
+  - from-scratch
+---
+
+# 🧠 Learn LLMs from First Principles — A Complete Study Plan
+
+> **Goal:** Go from zero to confidently understanding, building, fine-tuning, aligning, and deploying Large Language Models.
+>
+> **Time commitment:** ~20 weeks (10–15 hrs/week) — flexible, self-paced.
+>
+> **Prerequisites:** Python proficiency. Basic calculus & linear algebra (or willingness to learn in Week 1).
+
+---
+
+## Table of Contents
+
+- [Overview & Philosophy](#overview--philosophy)
+- [Phase 1: Mathematical & Neural Network Foundations (Weeks 1–3)](#phase-1-mathematical--neural-network-foundations-weeks-13)
+- [Phase 2: The Transformer Architecture from Scratch (Weeks 4–6)](#phase-2-the-transformer-architecture-from-scratch-weeks-46)
+- [Phase 3: Language Modeling & Pretraining (Weeks 7–9)](#phase-3-language-modeling--pretraining-weeks-79)
+- [Phase 4: The Hugging Face Ecosystem (Weeks 10–12)](#phase-4-the-hugging-face-ecosystem-weeks-1012)
+- [Phase 5: Fine-Tuning & Alignment (Weeks 13–16)](#phase-5-fine-tuning--alignment-weeks-1316)
+- [Phase 6: Advanced Topics & Capstone (Weeks 17–20)](#phase-6-advanced-topics--capstone-weeks-1720)
+- [Reading List: Landmark Papers](#reading-list-landmark-papers)
+- [All Resources & Links](#all-resources--links)
+- [Appendix: Glossary of Key Terms](#appendix-glossary-of-key-terms)
+
+---
+
+## Overview & Philosophy
+
+This plan is built on one principle: **understand by building**. Rather than memorizing APIs, you will:
+
+1. **Derive** the math (attention, backpropagation, loss functions)
+2. **Implement** each concept in raw PyTorch before using libraries
+3. **Train** real models on real data
+4. **Read** the original papers that introduced each idea
+
+The plan follows a bottom-up progression:
+
+```
+Math Foundations ──► Neural Networks ──► Transformers ──► Language Models
+      │                     │                  │                │
+      ▼                     ▼                  ▼                ▼
+  Linear Algebra       Backprop &          Self-Attention     GPT-2 from
+  Calculus             Gradient Descent    Multi-Head Attn    Scratch
+  Probability          MLP, RNN, LSTM     LayerNorm, FFN     Tokenization
+                                          Pos. Encoding      Pretraining
+
+──► Fine-Tuning ──► Alignment ──► Reasoning ──► Agents & Deployment
+         │               │             │                │
+         ▼               ▼             ▼                ▼
+     SFT, LoRA       RLHF, DPO     GRPO, R1       Tool-use, RAG
+     Chat Templates  Reward Models  Chain-of-Thought  Deployment
+     PEFT            PPO            Open R1           Quantization
+```
+
+---
+
+## Phase 1: Mathematical & Neural Network Foundations (Weeks 1–3)
+
+### 🎯 Goal
+Build intuition for the math that powers every neural network, then implement a neural network from scratch.
+
+---
+
+### Week 1: Linear Algebra, Calculus & Probability
+
+> *"If you don't understand the math, you're just calling functions."*
+
+#### Study Material
+
+| Resource | Topic | Time | Link |
+|----------|-------|------|------|
+| 3Blue1Brown: Essence of Linear Algebra | Vectors, matrices, transformations, eigenvectors | ~3.5 hrs (16 videos) | [YouTube Playlist](https://www.youtube.com/playlist?list=PLZHQObOWTQDPD3MizzM2xVFitgF8hE_ab) |
+| 3Blue1Brown: Essence of Calculus | Derivatives, integrals, chain rule | ~3 hrs (12 videos) | [YouTube Playlist](https://www.youtube.com/playlist?list=PLZHQObOWTQDMsr9K-rj53DwVRMYO3t5Yr) |
+| StatQuest: Probability Fundamentals | Bayes theorem, distributions, MLE | ~2 hrs | [YouTube Channel](https://www.youtube.com/@statquest) |
+
+#### Key Concepts to Master
+
+- [ ] Matrix multiplication — what it *means* geometrically (not just how to compute it)
+- [ ] Dot product — as projection and similarity measure
+- [ ] Softmax function — converting raw scores to probabilities
+- [ ] Chain rule — the foundation of backpropagation
+- [ ] Cross-entropy loss — measuring how wrong a probability distribution is
+- [ ] Maximum likelihood estimation — why we minimize negative log-likelihood
+
+#### 📝 Exercise
+Implement these in Python (no NumPy):
+```python
+def matmul(A, B): ...          # Matrix multiplication from scratch
+def softmax(x): ...            # Softmax function
+def cross_entropy(pred, target): ...  # Cross-entropy loss
+```
+
+---
+
+### Week 2: Neural Networks from Scratch
+
+#### Study Material
+
+| Resource | Topic | Time | Link |
+|----------|-------|------|------|
+| 3Blue1Brown: Neural Networks | What is a neural network, gradient descent, backprop | ~1 hr (4 videos) | [YouTube Playlist](https://www.youtube.com/playlist?list=PLZHQObOWTQDNU6R1_67000Dx_ZCJB-3pi) |
+| 3Blue1Brown: Attention in Transformers | Visual explanation of attention mechanism | ~45 min | [YouTube](https://www.youtube.com/watch?v=eMlx5fFNoYc) |
+| Karpathy: The spelled-out intro to neural networks and backpropagation | Build micrograd — a tiny autograd engine | ~2.5 hrs | [YouTube](https://www.youtube.com/watch?v=VMj-3S1tku0) |
+
+#### Key Concepts to Master
+
+- [ ] Forward pass — computing outputs layer by layer
+- [ ] Loss function — quantifying prediction error
+- [ ] Backward pass (backpropagation) — computing gradients via chain rule
+- [ ] Gradient descent — updating weights to minimize loss
+- [ ] Learning rate — step size for weight updates
+- [ ] Computational graph — tracking operations for automatic differentiation
+
+#### 🔨 Project: Build micrograd
+Follow Karpathy's video and build a complete autograd engine from scratch:
+- `Value` class with `__add__`, `__mul__`, `__pow__`
+- Automatic gradient computation via `.backward()`
+- Train a simple MLP on a toy dataset
+- **Repo:** [github.com/karpathy/micrograd](https://github.com/karpathy/micrograd)
+
+---
+
+### Week 3: PyTorch Fundamentals & MLPs
+
+#### Study Material
+
+| Resource | Topic | Time | Link |
+|----------|-------|------|------|
+| PyTorch Official: 60-min Blitz | Tensors, autograd, nn.Module | ~2 hrs | [pytorch.org/tutorials](https://pytorch.org/tutorials/beginner/deep_learning_60min_blitz.html) |
+| fast.ai Lesson 1 | Practical deep learning, training loops | ~2 hrs | [course.fast.ai](https://course.fast.ai/) |
+| Karpathy: makemore Part 1 (Bigrams) | Character-level language model, counting | ~1.5 hrs | [YouTube](https://www.youtube.com/watch?v=PaCmpygFfXo) |
+| Karpathy: makemore Part 2 (MLP) | MLP language model (Bengio et al., 2003) | ~1.5 hrs | [YouTube](https://www.youtube.com/watch?v=TCH_1BHY58I) |
+
+#### Key Concepts to Master
+
+- [ ] `torch.Tensor` — creation, indexing, broadcasting, GPU transfer
+- [ ] `torch.nn.Module` — defining layers, forward method
+- [ ] `torch.optim` — SGD, Adam
+- [ ] `loss.backward()` — automatic differentiation
+- [ ] Training loop: forward → loss → backward → step → zero_grad
+- [ ] Bigram model — simplest language model (predicting next character from previous one)
+- [ ] MLP language model — predicting next character from a window of previous characters
+
+#### 🔨 Project: Character-level Language Model
+Follow makemore Parts 1 & 2:
+1. Build a bigram model (lookup table / counting approach)
+2. Build an MLP model that takes N characters and predicts the next
+3. Train on a names dataset and generate new names
+- **Repo:** [github.com/karpathy/makemore](https://github.com/karpathy/makemore)
+
+---
+
+## Phase 2: The Transformer Architecture from Scratch (Weeks 4–6)
+
+### 🎯 Goal
+Understand every component of the Transformer and implement it from scratch in PyTorch.
+
+---
+
+### Week 4: Deep Dive into Training Dynamics
+
+#### Study Material
+
+| Resource | Topic | Time | Link |
+|----------|-------|------|------|
+| Karpathy: makemore Part 3 | Batch normalization, weight initialization, activation statistics | ~1.5 hrs | [YouTube](https://www.youtube.com/watch?v=P6sfmUTpUmc) |
+| Karpathy: makemore Part 4 | Becoming a backprop ninja — manual backpropagation | ~1.5 hrs | [YouTube](https://www.youtube.com/watch?v=q8SA3rM6ckI) |
+| Karpathy: makemore Part 5 | WaveNet-style deep networks, dilated causal convolutions | ~1 hr | [YouTube](https://www.youtube.com/watch?v=t3YJ5hKiMQ0) |
+
+#### Key Concepts to Master
+
+- [ ] Batch normalization — why it stabilizes training, running mean/var
+- [ ] Weight initialization — Kaiming/He init, why it matters
+- [ ] Activation statistics — dead neurons, saturation, vanishing gradients
+- [ ] Manual backpropagation — computing gradients by hand through every operation
+- [ ] Residual connections — why `x + f(x)` helps deep networks
+- [ ] Hierarchical models — building deeper architectures
+
+#### 📝 Exercise
+Take your makemore MLP and manually compute the gradient of every parameter by hand (no `.backward()`). Verify against PyTorch autograd.
+
+---
+
+### Week 5: The Transformer — Theory & Paper Reading
+
+#### Study Material
+
+| Resource | Topic | Time | Link |
+|----------|-------|------|------|
+| **Paper: "Attention Is All You Need"** | The original Transformer paper | ~3 hrs (multiple readings) | [arxiv:1706.03762](https://arxiv.org/abs/1706.03762) |
+| 3Blue1Brown: Attention in Transformers | Visual explanation of QKV, multi-head attention | ~45 min | [YouTube](https://www.youtube.com/watch?v=eMlx5fFNoYc) |
+| Jay Alammar: The Illustrated Transformer | Step-by-step visual walkthrough | ~1.5 hrs | [jalammar.github.io](https://jalammar.github.io/illustrated-transformer/) |
+| Lilian Weng: "Attention? Attention!" | Comprehensive attention mechanisms survey | ~2 hrs | [lilianweng.github.io](https://lilianweng.github.io/posts/2018-06-24-attention/) |
+
+#### Key Sections of "Attention Is All You Need" to Study
+
+| Section | What to Learn | Core Formula |
+|---------|---------------|--------------|
+| §3.1 Encoder & Decoder Stacks | N=6 layers, residual + LayerNorm | `LayerNorm(x + Sublayer(x))` |
+| §3.2.1 Scaled Dot-Product Attention | The heart of everything | `Attention(Q,K,V) = softmax(QK^T / √d_k) V` |
+| §3.2.2 Multi-Head Attention | Parallel attention in subspaces | `MultiHead = Concat(head_1..h) W^O` |
+| §3.3 Position-wise FFN | Two-layer MLP per position | `FFN(x) = max(0, xW₁+b₁)W₂+b₂` |
+| §3.5 Positional Encoding | Sinusoidal position signals | `PE(pos,2i) = sin(pos/10000^(2i/d))` |
+| §4 Why Self-Attention | O(1) sequential ops vs O(n) for RNN | Complexity comparison table |
+| §5.3 Optimizer | Learning rate warmup schedule | `lr = d^(-0.5) · min(step^(-0.5), step·warmup^(-1.5))` |
+
+#### Key Concepts to Master
+
+- [ ] Query, Key, Value — what each represents intuitively
+- [ ] Scaled dot-product attention — why we scale by `√d_k`
+- [ ] Multi-head attention — why multiple heads are better than one
+- [ ] Causal masking — preventing the model from "seeing the future"
+- [ ] Positional encoding — how the model knows word order
+- [ ] Layer normalization — stabilizing training in transformers
+- [ ] Encoder vs. Decoder vs. Encoder-Decoder architectures
+- [ ] Residual connections — enabling gradient flow in deep networks
+
+#### 📝 Exercise
+Draw the complete Transformer architecture from memory, labeling every component with its dimensions (for d_model=512, h=8, d_k=64, d_ff=2048, N=6).
+
+---
+
+### Week 6: Build GPT from Scratch
+
+> **This is the most important week in the entire plan.**
+
+#### Study Material
+
+| Resource | Topic | Time | Link |
+|----------|-------|------|------|
+| **Karpathy: "Let's Build GPT: from scratch, in code, spelled out"** | Complete GPT implementation in PyTorch | ~2 hrs | [YouTube](https://youtu.be/kCc8FmEb1nY) |
+| nanoGPT `model.py` | Production-quality GPT implementation (~300 lines) | ~3 hrs (read line by line) | [GitHub](https://github.com/karpathy/nanoGPT/blob/master/model.py) |
+| nanoGPT `train.py` | Training loop with DDP, AMP, gradient accumulation | ~2 hrs | [GitHub](https://github.com/karpathy/nanoGPT/blob/master/train.py) |
+
+#### What You'll Build (following the video)
+
+```
+Input Text
+    ↓
+[Token Embedding] + [Position Embedding]
+    ↓
+┌─────────────────────────────────┐
+│  Transformer Block (×N)         │
+│  ├── LayerNorm                  │
+│  ├── Multi-Head Self-Attention  │
+│  │   ├── Q = x @ W_q           │
+│  │   ├── K = x @ W_k           │
+│  │   ├── V = x @ W_v           │
+│  │   ├── Causal Mask            │
+│  │   └── softmax(QK^T/√d) @ V  │
+│  ├── Residual Connection        │
+│  ├── LayerNorm                  │
+│  ├── Feed-Forward (MLP)         │
+│  │   ├── Linear → GELU → Linear│
+│  └── Residual Connection        │
+└─────────────────────────────────┘
+    ↓
+[LayerNorm]
+    ↓
+[Linear → Logits]
+    ↓
+[Softmax → Next Token Probabilities]
+```
+
+#### Key Concepts to Master
+
+- [ ] Token embedding table — mapping integers to vectors
+- [ ] Position embedding — learned (GPT-2 style) vs sinusoidal
+- [ ] Self-attention with causal mask — the `tril` trick
+- [ ] Multi-head attention — splitting d_model into h heads
+- [ ] Feed-forward network — expanding and contracting dimensions
+- [ ] Residual stream — the "highway" through the network
+- [ ] Pre-norm vs post-norm — GPT-2 uses pre-norm
+- [ ] Weight tying — sharing embedding and output projection weights
+- [ ] Temperature and top-k sampling — controlling text generation
+
+#### 🔨 Project: Your Own nanoGPT
+1. Watch the video, pause, and implement each component yourself
+2. Train a character-level GPT on Shakespeare (~1M characters)
+3. Generate text and observe how quality improves with training
+4. Experiment: change number of heads, layers, embedding dimension
+5. **Deliverable:** A working GPT that generates coherent Shakespeare text
+
+---
+
+## Phase 3: Language Modeling & Pretraining (Weeks 7–9)
+
+### 🎯 Goal
+Understand tokenization, scaling laws, and the full pretraining pipeline.
+
+---
+
+### Week 7: Tokenization — The Unsung Hero
+
+#### Study Material
+
+| Resource | Topic | Time | Link |
+|----------|-------|------|------|
+| Karpathy: "Let's build the GPT Tokenizer" | BPE from scratch, tiktoken, sentencepiece | ~2 hrs | [YouTube](https://www.youtube.com/watch?v=zduSFxRajkE) |
+| HF NLP Course: Chapter 6 — Tokenizers | BPE, WordPiece, Unigram, `🤗 Tokenizers` library | ~3 hrs | [HF Course Ch.6](https://huggingface.co/learn/nlp-course/chapter6/1) |
+| **Paper: Sennrich et al. (2016)** | Original BPE for NMT | ~1 hr | [arxiv:1508.07909](https://arxiv.org/abs/1508.07909) |
+
+#### Key Concepts to Master
+
+- [ ] Why tokenization matters — it defines what the model "sees"
+- [ ] Character-level vs word-level vs subword tokenization
+- [ ] Byte Pair Encoding (BPE) — the greedy merge algorithm
+- [ ] WordPiece — BPE variant used by BERT
+- [ ] Unigram — probabilistic tokenizer used by T5
+- [ ] SentencePiece — language-agnostic tokenization
+- [ ] Special tokens — `<bos>`, `<eos>`, `<pad>`, `<unk>`
+- [ ] Vocabulary size tradeoffs — larger vocab = shorter sequences but bigger embedding table
+- [ ] The `🤗 Tokenizers` library — training custom tokenizers in Rust-speed
+
+#### 🔨 Project: Train Your Own BPE Tokenizer
+1. Implement BPE from scratch in Python (follow Karpathy)
+2. Train a `🤗 Tokenizers` BPE tokenizer on a custom corpus
+3. Compare vocabulary sizes: 1K, 10K, 50K — observe token lengths
+4. Decode some tokens and study what subwords the model learned
+
+---
+
+### Week 8: Scaling Laws & Pretraining Concepts
+
+#### Study Material
+
+| Resource | Topic | Time | Link |
+|----------|-------|------|------|
+| **Paper: "Scaling Laws for Neural Language Models" (Kaplan et al.)** | How loss scales with model size, data, compute | ~2 hrs | [arxiv:2001.08361](https://arxiv.org/abs/2001.08361) |
+| **Paper: "Training Compute-Optimal LLMs" (Chinchilla)** | Optimal ratio of model size to training tokens | ~2 hrs | [arxiv:2203.15556](https://arxiv.org/abs/2203.15556) |
+| **Paper: "Language Models are Few-Shot Learners" (GPT-3)** | In-context learning, few-shot prompting | ~3 hrs | [arxiv:2005.14165](https://arxiv.org/abs/2005.14165) |
+| Karpathy: "Reproducing GPT-2 (124M)" | Full pretraining pipeline on OpenWebText | ~4 hrs | [YouTube](https://www.youtube.com/watch?v=l8pRSuU81PU) |
+
+#### Key Concepts to Master
+
+- [ ] Scaling laws — `L(N) ∝ N^(-α)` — loss decreases as a power law of model size
+- [ ] Chinchilla-optimal — train N parameters on ~20N tokens
+- [ ] Compute-optimal training — how to budget FLOPs between model size and data
+- [ ] In-context learning — how large models learn from examples in the prompt
+- [ ] Emergent abilities — capabilities that appear only at scale
+- [ ] Data quality vs quantity — why curated data beats raw web scrapes
+- [ ] Mixed-precision training (fp16/bf16) — training faster with lower precision
+- [ ] Gradient accumulation — simulating large batches on small GPUs
+- [ ] Distributed Data Parallel (DDP) — training across multiple GPUs
+
+#### 📝 Exercise
+Calculate:
+1. How many FLOPs to train a 1B parameter model on 20B tokens?
+   - Rule of thumb: `FLOPs ≈ 6 × N × D` (N=params, D=tokens)
+2. If you have 8× A100 GPUs at 312 TFLOPS each, how long would training take?
+3. What's the Chinchilla-optimal model size for a 1T token dataset?
+
+---
+
+### Week 9: Pretraining a Small Model End-to-End
+
+#### Study Material
+
+| Resource | Topic | Time | Link |
+|----------|-------|------|------|
+| HF NLP Course: Chapter 7 — Main NLP Tasks | Causal LM pretraining with HF Transformers | ~4 hrs | [HF Course Ch.7](https://huggingface.co/learn/nlp-course/chapter7/1) |
+| nanoGPT README & `data/openwebtext/prepare.py` | Data preparation for pretraining | ~1 hr | [GitHub](https://github.com/karpathy/nanoGPT) |
+| **Paper: SmolLM2 (Allal et al.)** | How to train a good small LM | ~2 hrs | [arxiv:2502.02737](https://arxiv.org/abs/2502.02737) |
+
+#### Key Concepts to Master
+
+- [ ] Dataset preparation — downloading, cleaning, tokenizing, chunking
+- [ ] Data loading — efficient batching, shuffling, packing sequences
+- [ ] Learning rate schedule — warmup + cosine decay
+- [ ] Loss curves — what a healthy training run looks like
+- [ ] Evaluation — perplexity, held-out validation loss
+- [ ] Checkpointing — saving model state during training
+- [ ] The `Trainer` API — HF's high-level training abstraction
+
+#### 🔨 Project: Pretrain a 10M–50M Parameter GPT
+Using nanoGPT or HF Transformers:
+1. Prepare a dataset (TinyStories, OpenWebText subset, or custom)
+2. Configure a small GPT (6–12 layers, 256–512 dim, 4–8 heads)
+3. Train for ~10K steps on a single GPU
+4. Plot the loss curve, compute perplexity
+5. Generate text samples at different checkpoints — watch quality improve
+6. Push the model to [Hugging Face Hub](https://huggingface.co)
+
+---
+
+## Phase 4: The Hugging Face Ecosystem (Weeks 10–12)
+
+### 🎯 Goal
+Master the tools you'll use daily: Transformers, Datasets, Hub, and Gradio.
+
+---
+
+### Week 10: Transformers Library & the Hub
+
+#### Study Material
+
+| Resource | Topic | Time | Link |
+|----------|-------|------|------|
+| HF NLP Course: Chapter 1 — Transformer Models | `pipeline()`, model architectures, use cases | ~2 hrs | [HF Course Ch.1](https://huggingface.co/learn/nlp-course/chapter1/1) |
+| HF NLP Course: Chapter 2 — Using Transformers | `AutoModel`, `AutoTokenizer`, batching | ~3 hrs | [HF Course Ch.2](https://huggingface.co/learn/nlp-course/chapter2/1) |
+| HF NLP Course: Chapter 4 — Sharing on the Hub | Model cards, `push_to_hub()`, versioning | ~1.5 hrs | [HF Course Ch.4](https://huggingface.co/learn/nlp-course/chapter4/1) |
+
+#### Key Concepts to Master
+
+- [ ] `pipeline()` — high-level inference in one line
+- [ ] `AutoModel.from_pretrained()` — loading any model architecture
+- [ ] `AutoTokenizer.from_pretrained()` — loading the matching tokenizer
+- [ ] Encoder models (BERT) vs Decoder models (GPT) vs Encoder-Decoder (T5)
+- [ ] Model cards — documenting your models for others
+- [ ] The Hub — versioning, repos, community features
+
+#### 🔨 Mini-Project
+Load 5 different models from the Hub and run inference:
+1. Text generation (GPT-2 / SmolLM2)
+2. Text classification (BERT / DistilBERT)
+3. Summarization (T5 / BART)
+4. Translation (MarianMT)
+5. Fill-mask (BERT)
+
+---
+
+### Week 11: Datasets, Fine-Tuning & Debugging
+
+#### Study Material
+
+| Resource | Topic | Time | Link |
+|----------|-------|------|------|
+| HF NLP Course: Chapter 3 — Fine-Tuning | `Trainer` API, Accelerate, custom training loops | ~4 hrs | [HF Course Ch.3](https://huggingface.co/learn/nlp-course/chapter3/1) |
+| HF NLP Course: Chapter 5 — Datasets Library | Loading, streaming, Arrow format, processing | ~3 hrs | [HF Course Ch.5](https://huggingface.co/learn/nlp-course/chapter5/1) |
+| HF NLP Course: Chapter 8 — Debugging | Common training failures, community help | ~1.5 hrs | [HF Course Ch.8](https://huggingface.co/learn/nlp-course/chapter8/1) |
+
+#### Key Concepts to Master
+
+- [ ] `datasets.load_dataset()` — loading from Hub, local files, streaming
+- [ ] `dataset.map()` — applying transformations efficiently
+- [ ] Apache Arrow format — why HF datasets are fast
+- [ ] `Trainer` API — `TrainingArguments`, callbacks, logging
+- [ ] Custom training loop with `Accelerate`
+- [ ] Common errors: shape mismatches, tokenizer/model mismatches, OOM
+- [ ] Gradient checkpointing — trading compute for memory
+
+#### 🔨 Project: Fine-Tune a Text Classifier
+1. Load a sentiment dataset (e.g., `stanfordnlp/imdb`)
+2. Tokenize and prepare data with `datasets`
+3. Fine-tune `distilbert-base-uncased` with `Trainer`
+4. Evaluate accuracy on test set
+5. Push the fine-tuned model to the Hub
+6. Build a Gradio demo (preview of next week)
+
+---
+
+### Week 12: Demos, Data Annotation & the Full Workflow
+
+#### Study Material
+
+| Resource | Topic | Time | Link |
+|----------|-------|------|------|
+| HF NLP Course: Chapter 9 — Gradio | Building interactive ML demos | ~2 hrs | [HF Course Ch.9](https://huggingface.co/learn/nlp-course/chapter9/1) |
+| HF NLP Course: Chapter 10 — Data Annotation | Argilla for dataset curation | ~2 hrs | [HF Course Ch.10](https://huggingface.co/learn/nlp-course/chapter10/1) |
+| HF Spaces Documentation | Deploying models as web apps | ~1 hr | [HF Spaces](https://huggingface.co/docs/hub/spaces) |
+
+#### Key Concepts to Master
+
+- [ ] `gr.Interface()` — creating demos with Gradio
+- [ ] `gr.Blocks()` — advanced layouts and interactivity
+- [ ] Hugging Face Spaces — deploying demos for free
+- [ ] Data annotation with Argilla — creating high-quality datasets
+- [ ] Human-in-the-loop workflows — iterating on data quality
+
+#### 🔨 Project: Deploy a Model on HF Spaces
+1. Take your fine-tuned classifier from Week 11
+2. Build a Gradio app with text input → sentiment prediction
+3. Deploy on Hugging Face Spaces
+4. Share the link and get feedback
+
+---
+
+## Phase 5: Fine-Tuning & Alignment (Weeks 13–16)
+
+### 🎯 Goal
+Learn the modern LLM training stack: SFT, LoRA, DPO, RLHF, and GRPO.
+
+---
+
+### Week 13: Supervised Fine-Tuning (SFT) & Chat Models
+
+#### Study Material
+
+| Resource | Topic | Time | Link |
+|----------|-------|------|------|
+| HF NLP Course: Chapter 11 — SFT | Chat templates, `SFTTrainer`, evaluation | ~4 hrs | [HF Course Ch.11](https://huggingface.co/learn/nlp-course/chapter11/1) |
+| smol-course: Module 1 — Instruction Tuning | Hands-on SFT with SmolLM2 | ~3 hrs | [GitHub](https://github.com/huggingface/smol-course) |
+| **Paper: "Training language models to follow instructions" (InstructGPT)** | The paper that started it all (OpenAI, 2022) | ~2 hrs | [arxiv:2203.02155](https://arxiv.org/abs/2203.02155) |
+
+#### Key Concepts to Master
+
+- [ ] Chat templates (ChatML format) — `system`, `user`, `assistant` roles
+- [ ] The `messages` format — structured conversation data
+- [ ] `SFTTrainer` from TRL — the standard fine-tuning trainer
+- [ ] Dataset preparation — converting raw text to ChatML format
+- [ ] Packing — fitting multiple short examples in one sequence
+- [ ] Evaluation — loss curves, manual quality checks, benchmarks
+- [ ] Base model → Instruct model transformation
+
+#### Dataset Format for SFT
+```json
+{
+  "messages": [
+    {"role": "system", "content": "You are a helpful assistant."},
+    {"role": "user", "content": "What is the capital of France?"},
+    {"role": "assistant", "content": "The capital of France is Paris."}
+  ]
+}
+```
+
+#### 🔨 Project: Fine-Tune SmolLM2 into a Chat Model
+1. Load [HuggingFaceTB/SmolLM2-135M](https://huggingface.co/HuggingFaceTB/SmolLM2-135M)
+2. Prepare data from [HuggingFaceTB/smoltalk](https://huggingface.co/datasets/HuggingFaceTB/smoltalk)
+3. Apply chat template and train with `SFTTrainer`
+4. Chat with your model and evaluate quality
+5. Push to Hub as `your-username/SmolLM2-135M-Chat`
+
+---
+
+### Week 14: Parameter-Efficient Fine-Tuning (LoRA / QLoRA)
+
+#### Study Material
+
+| Resource | Topic | Time | Link |
+|----------|-------|------|------|
+| smol-course: Module 3 — PEFT | LoRA, QLoRA, adapter merging | ~3 hrs | [GitHub](https://github.com/huggingface/smol-course) |
+| **Paper: "LoRA: Low-Rank Adaptation of Large Language Models"** | The LoRA method | ~2 hrs | [arxiv:2106.09685](https://arxiv.org/abs/2106.09685) — [HF Paper](https://hf.co/papers/2106.09685) |
+| HF PEFT Documentation | `LoraConfig`, `get_peft_model`, merging | ~2 hrs | [HF PEFT Docs](https://huggingface.co/docs/peft) |
+
+#### Key Concepts to Master
+
+- [ ] Why full fine-tuning is expensive — every parameter needs gradients + optimizer states
+- [ ] Low-rank decomposition — `W + ΔW = W + BA` where B is (d×r) and A is (r×d)
+- [ ] Rank `r` — the bottleneck dimension (typical: 8, 16, 32, 64)
+- [ ] `lora_alpha` — scaling factor for LoRA updates
+- [ ] Target modules — which layers to add LoRA to (`q_proj`, `v_proj`, `k_proj`, etc.)
+- [ ] QLoRA — LoRA on 4-bit quantized models (fits 7B models on consumer GPUs!)
+- [ ] Adapter merging — combining LoRA weights back into the base model
+- [ ] Memory savings — LoRA trains ~0.1–1% of parameters
+
+#### The Math of LoRA
+```
+Original: h = Wx                   (d × d matrix, d² parameters)
+LoRA:     h = Wx + BAx             (B: d×r, A: r×d, only 2dr parameters)
+                                    With r=16, d=4096: 131K vs 16.7M params (125× reduction)
+```
+
+#### 🔨 Project: QLoRA Fine-Tune a 1.7B Model on Consumer GPU
+1. Load [HuggingFaceTB/SmolLM2-1.7B](https://huggingface.co/HuggingFaceTB/SmolLM2-1.7B) in 4-bit
+2. Add LoRA adapters with `peft`
+3. Train on a domain-specific dataset (e.g., coding, medical, legal)
+4. Compare generation quality: base vs LoRA-tuned
+5. Merge adapters and push the merged model to Hub
+
+---
+
+### Week 15: Preference Alignment — DPO & RLHF
+
+#### Study Material
+
+| Resource | Topic | Time | Link |
+|----------|-------|------|------|
+| smol-course: Module 2 — Preference Alignment | DPO training hands-on | ~3 hrs | [GitHub](https://github.com/huggingface/smol-course) |
+| **Paper: "Direct Preference Optimization" (Rafailov et al.)** | DPO — RLHF without a reward model | ~2 hrs | [arxiv:2305.18290](https://arxiv.org/abs/2305.18290) |
+| **Paper: "Training language models to follow instructions" (InstructGPT)** | Original RLHF pipeline | ~2 hrs | [arxiv:2203.02155](https://arxiv.org/abs/2203.02155) |
+| TRL DPO Documentation | `DPOTrainer`, `DPOConfig` | ~1 hr | [HF TRL Docs](https://huggingface.co/docs/trl/dpo_trainer) |
+
+#### Key Concepts to Master
+
+- [ ] The alignment problem — why SFT alone isn't enough
+- [ ] Human preferences — "which response is better?" → preference data
+- [ ] RLHF pipeline: SFT → Reward Model → PPO optimization
+- [ ] DPO — bypasses the reward model, optimizes preferences directly
+- [ ] The DPO loss: `L_DPO = -log σ(β · (log π(y_w|x)/π_ref(y_w|x) - log π(y_l|x)/π_ref(y_l|x)))`
+- [ ] `β` parameter — controls deviation from reference policy
+- [ ] Preference dataset format — `prompt`, `chosen`, `rejected` columns
+- [ ] KL divergence — preventing the model from straying too far
+
+#### Dataset Format for DPO
+```json
+{
+  "prompt": "Explain quantum computing simply.",
+  "chosen": "Quantum computing uses qubits that can be 0, 1, or both at once...",
+  "rejected": "Quantum computing is a type of computing that uses quantum mechanics..."
+}
+```
+
+#### 🔨 Project: Align a Model with DPO
+1. Start from your SFT model (Week 13)
+2. Load a preference dataset
+3. Train with `DPOTrainer`
+4. Compare: base → SFT → DPO outputs side by side
+5. Evaluate helpfulness and safety improvements
+
+---
+
+### Week 16: Reasoning Models — GRPO & Open R1
+
+#### Study Material
+
+| Resource | Topic | Time | Link |
+|----------|-------|------|------|
+| HF NLP Course: Chapter 12 — Reasoning Models | GRPO, DeepSeek R1, Open R1 project | ~4 hrs | [HF Course Ch.12](https://huggingface.co/learn/nlp-course/chapter12/1) |
+| **Paper: "DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning"** | The breakthrough reasoning model | ~3 hrs | [arxiv:2501.12948](https://arxiv.org/abs/2501.12948) — [HF Paper](https://hf.co/papers/2501.12948) |
+| Deep RL Course: Units 1–4 | RL fundamentals, policy gradients, PPO | ~6 hrs | [HF Deep RL Course](https://huggingface.co/learn/deep-rl-course/unit0/introduction) |
+
+#### Key Concepts to Master
+
+- [ ] Why RL for reasoning — SFT on reasoning traces vs discovering reasoning via RL
+- [ ] DeepSeek-R1-Zero — reasoning emerges purely from RL (no SFT)
+- [ ] Group Relative Policy Optimization (GRPO) — reward relative to group average
+- [ ] Reward functions — correctness checkers, format validators
+- [ ] Chain-of-thought — step-by-step reasoning in `<think>` tags
+- [ ] Cold-start data — bootstrapping reasoning with SFT before RL
+- [ ] Multi-stage training: SFT → RL (reasoning) → SFT (readability) → RL (all tasks)
+- [ ] `GRPOTrainer` from TRL — practical implementation
+
+#### The GRPO Algorithm (Simplified)
+```
+For each prompt x:
+  1. Generate K responses: {y₁, y₂, ..., yₖ}
+  2. Score each: {r₁, r₂, ..., rₖ}
+  3. Compute group advantage: Aᵢ = (rᵢ - mean(r)) / std(r)
+  4. Update policy to increase probability of high-advantage responses
+  5. Apply KL penalty to stay close to reference model
+```
+
+#### 🔨 Project: Train a Math Reasoning Model with GRPO
+1. Start from a small instruct model
+2. Prepare a math dataset with verifiable answers (GSM8K format)
+3. Define reward functions (correctness + format)
+4. Train with `GRPOTrainer`
+5. Evaluate on held-out math problems
+6. Observe the model learn to use `<think>` reasoning
+
+---
+
+## Phase 6: Advanced Topics & Capstone (Weeks 17–20)
+
+### 🎯 Goal
+Evaluation, synthetic data, deployment, agents, and a capstone project.
+
+---
+
+### Week 17: Evaluation & Benchmarks
+
+#### Study Material
+
+| Resource | Topic | Time | Link |
+|----------|-------|------|------|
+| smol-course: Module 4 — Evaluation | LLM evaluation, `lighteval` | ~3 hrs | [GitHub](https://github.com/huggingface/smol-course) |
+| HF `lighteval` Documentation | Running standardized benchmarks | ~2 hrs | [HF lighteval Docs](https://huggingface.co/docs/lighteval) |
+| Open LLM Leaderboard | Understanding LLM benchmarks | ~1 hr | [HF Leaderboard](https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard) |
+
+#### Key Concepts to Master
+
+- [ ] Perplexity — `exp(average cross-entropy loss)` — lower is better
+- [ ] MMLU — Massive Multitask Language Understanding (knowledge test)
+- [ ] GSM8K — Grade School Math 8K (reasoning benchmark)
+- [ ] HumanEval — coding benchmark (pass@k)
+- [ ] TruthfulQA — measuring hallucination tendencies
+- [ ] LLM-as-judge — using a strong LLM to evaluate a weaker one
+- [ ] Contamination — when evaluation data leaks into training data
+
+#### 🔨 Project: Benchmark Your Models
+Run `lighteval` on all models you've trained so far. Create a comparison table:
+| Model | Perplexity | MMLU | GSM8K | Notes |
+|-------|-----------|------|-------|-------|
+| Base SmolLM2 | ... | ... | ... | Pretrained |
+| + SFT | ... | ... | ... | Week 13 |
+| + LoRA | ... | ... | ... | Week 14 |
+| + DPO | ... | ... | ... | Week 15 |
+| + GRPO | ... | ... | ... | Week 16 |
+
+---
+
+### Week 18: Synthetic Data & Inference Optimization
+
+#### Study Material
+
+| Resource | Topic | Time | Link |
+|----------|-------|------|------|
+| smol-course: Module 6 — Synthetic Data | Generating training data with LLMs | ~3 hrs | [GitHub](https://github.com/huggingface/smol-course) |
+| smol-course: Module 7 — Inference | Quantization, vLLM, optimization | ~3 hrs | [GitHub](https://github.com/huggingface/smol-course) |
+| HF `bitsandbytes` Documentation | 4-bit and 8-bit quantization | ~1 hr | [HF BnB Docs](https://huggingface.co/docs/bitsandbytes) |
+| HF `distilabel` Documentation | Synthetic data pipelines | ~2 hrs | [HF Distilabel Docs](https://huggingface.co/docs/distilabel) |
+
+#### Key Concepts to Master
+
+- [ ] Knowledge distillation — training a small model on a large model's outputs
+- [ ] Synthetic data generation — using LLMs to create training data
+- [ ] Data quality filtering — removing bad synthetic examples
+- [ ] Quantization (INT8, INT4, GPTQ, AWQ) — reducing model size
+- [ ] KV-cache — speeding up autoregressive generation
+- [ ] Batched inference — processing multiple requests efficiently
+- [ ] vLLM — high-throughput inference server
+- [ ] Speculative decoding — using a small model to speed up a large model
+
+---
+
+### Week 19: AI Agents & Tool Use
+
+#### Study Material
+
+| Resource | Topic | Time | Link |
+|----------|-------|------|------|
+| HF Agents Course: Units 0–2 | Agent fundamentals, frameworks | ~6 hrs | [HF Agents Course](https://huggingface.co/learn/agents-course/unit0/introduction) |
+| smol-course: Module 8 — Agents | Building agents with smolagents | ~3 hrs | [GitHub](https://github.com/huggingface/smol-course) |
+| HF `smolagents` Documentation | Lightweight agent framework | ~2 hrs | [HF smolagents Docs](https://huggingface.co/docs/smolagents) |
+
+#### Key Concepts to Master
+
+- [ ] Agent loop: Thought → Action → Observation → repeat
+- [ ] Tool definition — giving LLMs access to functions
+- [ ] Function calling format — how models invoke tools
+- [ ] ReAct pattern — reasoning + acting in alternation
+- [ ] RAG (Retrieval Augmented Generation) — grounding in external knowledge
+- [ ] Code agents — generating and executing Python code
+- [ ] Multi-agent systems — agents collaborating or delegating
+
+#### 🔨 Project: Build a RAG Agent
+1. Create a knowledge base from a set of documents
+2. Build a retrieval tool using sentence embeddings
+3. Wire up a smolagents agent that answers questions using retrieval
+4. Deploy as a Gradio app on HF Spaces
+
+---
+
+### Week 20: Capstone Project 🎓
+
+#### Choose One (or combine multiple):
+
+**Option A: Train and Deploy a Domain Expert Model**
+1. Collect/curate a domain-specific dataset (medical, legal, code, science)
+2. SFT a base model on your dataset
+3. Apply LoRA for efficient training
+4. Evaluate on domain-specific benchmarks
+5. Deploy as an API + Gradio demo on HF Spaces
+6. Write a detailed model card
+
+**Option B: Build a Reasoning Model for Math**
+1. Generate synthetic math reasoning data
+2. SFT a model on chain-of-thought examples
+3. Apply GRPO to improve reasoning quality
+4. Evaluate on GSM8K and MATH benchmarks
+5. Compare SFT-only vs SFT+GRPO
+6. Write a blog post about your findings
+
+**Option C: Create an End-to-End Agent System**
+1. Fine-tune a model for function calling
+2. Build a multi-tool agent (web search, calculator, code execution)
+3. Add RAG with a custom knowledge base
+4. Evaluate agent performance on tasks
+5. Deploy as a Space with observability/logging
+6. Write documentation for others to use
+
+#### Deliverables
+- [ ] Trained model(s) on Hugging Face Hub
+- [ ] Model card with training details, evaluations, limitations
+- [ ] Gradio demo deployed on HF Spaces
+- [ ] Blog post / README documenting your journey
+
+---
+
+## Reading List: Landmark Papers
+
+These are the papers that defined the field. Read them in this order as you progress through the plan.
+
+### Tier 1: Must Read (referenced directly in the plan)
+
+| # | Paper | Year | Key Contribution | Link |
+|---|-------|------|-------------------|------|
+| 1 | **Attention Is All You Need** (Vaswani et al.) | 2017 | Transformer architecture | [arxiv:1706.03762](https://arxiv.org/abs/1706.03762) |
+| 2 | **BERT** (Devlin et al.) | 2018 | Bidirectional pretraining, MLM | [arxiv:1810.04805](https://arxiv.org/abs/1810.04805) |
+| 3 | **Language Models are Few-Shot Learners** (GPT-3) | 2020 | In-context learning, scaling | [arxiv:2005.14165](https://arxiv.org/abs/2005.14165) |
+| 4 | **Training Compute-Optimal LLMs** (Chinchilla) | 2022 | Scaling laws, data vs params | [arxiv:2203.15556](https://arxiv.org/abs/2203.15556) |
+| 5 | **Training LMs to Follow Instructions** (InstructGPT) | 2022 | RLHF pipeline | [arxiv:2203.02155](https://arxiv.org/abs/2203.02155) |
+| 6 | **LoRA** (Hu et al.) | 2021 | Parameter-efficient fine-tuning | [arxiv:2106.09685](https://arxiv.org/abs/2106.09685) |
+| 7 | **Direct Preference Optimization** (Rafailov et al.) | 2023 | DPO — alignment without RL | [arxiv:2305.18290](https://arxiv.org/abs/2305.18290) |
+| 8 | **DeepSeek-R1** (DeepSeek AI) | 2025 | GRPO, reasoning via RL | [arxiv:2501.12948](https://arxiv.org/abs/2501.12948) |
+
+### Tier 2: Highly Recommended (deepens understanding)
+
+| # | Paper | Year | Key Contribution | Link |
+|---|-------|------|-------------------|------|
+| 9 | **Neural Machine Translation by Jointly Learning to Align and Translate** (Bahdanau et al.) | 2014 | Attention mechanism (before Transformers) | [arxiv:1409.0473](https://arxiv.org/abs/1409.0473) |
+| 10 | **A Neural Probabilistic Language Model** (Bengio et al.) | 2003 | Word embeddings, neural LMs | [jmlr.org](https://www.jmlr.org/papers/volume3/bengio03a/bengio03a.pdf) |
+| 11 | **LLaMA: Open and Efficient Foundation Language Models** (Touvron et al.) | 2023 | Efficient open-source LLM recipe | [arxiv:2302.13971](https://arxiv.org/abs/2302.13971) |
+| 12 | **QLoRA** (Dettmers et al.) | 2023 | 4-bit fine-tuning on consumer GPUs | [arxiv:2305.14314](https://arxiv.org/abs/2305.14314) |
+| 13 | **SmolLM2** (Allal et al.) | 2025 | Small model training recipe | [arxiv:2502.02737](https://arxiv.org/abs/2502.02737) |
+| 14 | **Chain-of-Thought Prompting** (Wei et al.) | 2022 | Step-by-step reasoning | [arxiv:2201.11903](https://arxiv.org/abs/2201.11903) |
+
+### Tier 3: Reference (consult when needed)
+
+| # | Paper | Year | Key Contribution | Link |
+|---|-------|------|-------------------|------|
+| 15 | **FlashAttention** (Dao et al.) | 2022 | Memory-efficient attention | [arxiv:2205.14135](https://arxiv.org/abs/2205.14135) |
+| 16 | **RoFormer / RoPE** (Su et al.) | 2021 | Rotary position embeddings | [arxiv:2104.09864](https://arxiv.org/abs/2104.09864) |
+| 17 | **GQA: Grouped Query Attention** (Ainslie et al.) | 2023 | Efficient attention variant | [arxiv:2305.13245](https://arxiv.org/abs/2305.13245) |
+| 18 | **Scaling Laws for Neural LMs** (Kaplan et al.) | 2020 | Power-law scaling | [arxiv:2001.08361](https://arxiv.org/abs/2001.08361) |
+
+---
+
+## All Resources & Links
+
+### 📚 Courses
+
+| Course | Provider | URL |
+|--------|----------|-----|
+| NLP / LLM Course (12 chapters) | Hugging Face | [huggingface.co/learn/nlp-course](https://huggingface.co/learn/nlp-course/chapter1/1) |
+| smol-course (8 modules) | Hugging Face | [github.com/huggingface/smol-course](https://github.com/huggingface/smol-course) |
+| Deep RL Course | Hugging Face | [huggingface.co/learn/deep-rl-course](https://huggingface.co/learn/deep-rl-course/unit0/introduction) |
+| AI Agents Course | Hugging Face | [huggingface.co/learn/agents-course](https://huggingface.co/learn/agents-course/unit0/introduction) |
+| Neural Networks: Zero to Hero | Andrej Karpathy | [YouTube Playlist](https://www.youtube.com/playlist?list=PLAqhIrjkxbuWI23v9cThsA9GvCAUhRvKZ) |
+| Practical Deep Learning | fast.ai | [course.fast.ai](https://course.fast.ai/) |
+
+### 🎬 Key Videos
+
+| Video | Creator | Duration | Link |
+|-------|---------|----------|------|
+| Let's Build GPT | Karpathy | ~2 hrs | [youtu.be/kCc8FmEb1nY](https://youtu.be/kCc8FmEb1nY) |
+| Let's Build the GPT Tokenizer | Karpathy | ~2 hrs | [YouTube](https://www.youtube.com/watch?v=zduSFxRajkE) |
+| Reproducing GPT-2 (124M) | Karpathy | ~4 hrs | [YouTube](https://www.youtube.com/watch?v=l8pRSuU81PU) |
+| Attention in Transformers | 3Blue1Brown | ~45 min | [YouTube](https://www.youtube.com/watch?v=eMlx5fFNoYc) |
+| Essence of Linear Algebra | 3Blue1Brown | ~3.5 hrs | [YouTube Playlist](https://www.youtube.com/playlist?list=PLZHQObOWTQDPD3MizzM2xVFitgF8hE_ab) |
+
+### 🛠️ Code Repositories
+
+| Repo | Description | Link |
+|------|-------------|------|
+| nanoGPT | Simplest GPT training code | [github.com/karpathy/nanoGPT](https://github.com/karpathy/nanoGPT) |
+| minGPT | Educational GPT implementation | [github.com/karpathy/minGPT](https://github.com/karpathy/minGPT) |
+| makemore | Character LM progression | [github.com/karpathy/makemore](https://github.com/karpathy/makemore) |
+| micrograd | Autograd engine from scratch | [github.com/karpathy/micrograd](https://github.com/karpathy/micrograd) |
+
+### 📦 HF Libraries
+
+| Library | Purpose | Docs |
+|---------|---------|------|
+| `transformers` | Model loading, inference, training | [huggingface.co/docs/transformers](https://huggingface.co/docs/transformers) |
+| `datasets` | Data loading & processing | [huggingface.co/docs/datasets](https://huggingface.co/docs/datasets) |
+| `tokenizers` | Fast tokenizer training | [huggingface.co/docs/tokenizers](https://huggingface.co/docs/tokenizers) |
+| `trl` | SFT, DPO, GRPO trainers | [huggingface.co/docs/trl](https://huggingface.co/docs/trl) |
+| `peft` | LoRA, QLoRA, adapters | [huggingface.co/docs/peft](https://huggingface.co/docs/peft) |
+| `accelerate` | Distributed & mixed-precision training | [huggingface.co/docs/accelerate](https://huggingface.co/docs/accelerate) |
+| `evaluate` | Metrics & evaluation | [huggingface.co/docs/evaluate](https://huggingface.co/docs/evaluate) |
+| `lighteval` | LLM benchmarking | [huggingface.co/docs/lighteval](https://huggingface.co/docs/lighteval) |
+| `smolagents` | Agent framework | [huggingface.co/docs/smolagents](https://huggingface.co/docs/smolagents) |
+| `gradio` | ML demos | [gradio.app](https://www.gradio.app/) |
+| `bitsandbytes` | Quantization | [huggingface.co/docs/bitsandbytes](https://huggingface.co/docs/bitsandbytes) |
+| `distilabel` | Synthetic data | [huggingface.co/docs/distilabel](https://huggingface.co/docs/distilabel) |
+
+### 🤖 Models to Use
+
+| Model | Size | Use Case | Link |
+|-------|------|----------|------|
+| SmolLM2-135M | 135M | Learning, fast experiments | [HuggingFaceTB/SmolLM2-135M](https://huggingface.co/HuggingFaceTB/SmolLM2-135M) |
+| SmolLM2-360M | 360M | Small model training | [HuggingFaceTB/SmolLM2-360M](https://huggingface.co/HuggingFaceTB/SmolLM2-360M) |
+| SmolLM2-1.7B | 1.7B | QLoRA on consumer GPU | [HuggingFaceTB/SmolLM2-1.7B](https://huggingface.co/HuggingFaceTB/SmolLM2-1.7B) |
+| GPT-2 | 124M | Reference implementation | [openai-community/gpt2](https://huggingface.co/openai-community/gpt2) |
+| DistilBERT | 66M | Classification tasks | [distilbert-base-uncased](https://huggingface.co/distilbert-base-uncased) |
+
+### 📊 Datasets to Use
+
+| Dataset | Format | Use Case | Link |
+|---------|--------|----------|------|
+| smoltalk | ChatML (`messages`) | SFT training | [HuggingFaceTB/smoltalk](https://huggingface.co/datasets/HuggingFaceTB/smoltalk) |
+| Tiny Shakespeare | Raw text | nanoGPT pretraining | Included in nanoGPT repo |
+| TinyStories | Text | Small model pretraining | [roneneldan/TinyStories](https://huggingface.co/datasets/roneneldan/TinyStories) |
+| IMDb | Classification | Fine-tuning practice | [stanfordnlp/imdb](https://huggingface.co/datasets/stanfordnlp/imdb) |
+| GSM8K | Math QA | Reasoning evaluation | [openai/gsm8k](https://huggingface.co/datasets/openai/gsm8k) |
+
+---
+
+## Appendix: Glossary of Key Terms
+
+| Term | Definition |
+|------|-----------|
+| **Attention** | Mechanism that lets each token attend to all other tokens, computing relevance-weighted representations. Core formula: `softmax(QK^T/√d_k)V` |
+| **Autoregressive** | Generating one token at a time, each conditioned on all previous tokens. GPT-style models. |
+| **Backpropagation** | Algorithm for computing gradients of the loss w.r.t. all parameters by applying the chain rule backward through the computation graph. |
+| **BPE (Byte Pair Encoding)** | Tokenization algorithm that iteratively merges the most frequent pair of tokens. Used by GPT-2, GPT-3, LLaMA. |
+| **Causal Mask** | Lower-triangular mask that prevents tokens from attending to future positions. Makes the model autoregressive. |
+| **ChatML** | Standard format for chat data: list of `{role, content}` dictionaries with roles `system`, `user`, `assistant`. |
+| **Cross-Entropy Loss** | Standard loss for classification/language modeling: `-Σ yᵢ log(ŷᵢ)`. Measures how well predicted distribution matches target. |
+| **DPO (Direct Preference Optimization)** | Alignment method that directly optimizes the policy from preference pairs, without training a separate reward model. |
+| **Embedding** | Dense vector representation of a discrete token. Learned lookup table mapping token IDs to vectors. |
+| **Fine-Tuning** | Continuing training of a pretrained model on a specific downstream task or dataset. |
+| **GRPO (Group Relative Policy Optimization)** | RL algorithm that updates the policy based on relative advantage within a group of sampled responses. Used by DeepSeek-R1. |
+| **Gradient Accumulation** | Simulating large batch sizes by accumulating gradients over multiple forward/backward passes before updating weights. |
+| **KV-Cache** | Caching key and value tensors from previous tokens during autoregressive generation, avoiding recomputation. |
+| **Layer Normalization** | Normalizing activations across the feature dimension (not the batch dimension). Stabilizes Transformer training. |
+| **LoRA** | Adding small low-rank matrices (B×A where rank r << d) to existing weight matrices. Trains ~0.1% of parameters. |
+| **Perplexity** | `exp(cross-entropy loss)`. Intuitively: how many tokens the model is "confused" between. Lower = better. |
+| **Positional Encoding** | Information added to token embeddings so the model knows the order of tokens. Sinusoidal (original) or learned (GPT-2). |
+| **Pretraining** | Initial training on a large unlabeled corpus (next-token prediction). Creates the base model. |
+| **QLoRA** | LoRA applied to a 4-bit quantized base model. Enables fine-tuning 65B models on a single 48GB GPU. |
+| **Quantization** | Reducing numerical precision (fp32 → fp16 → int8 → int4) to reduce model size and speed up inference. |
+| **Residual Connection** | `output = x + f(x)`. Allows gradients to flow directly through the network, enabling very deep models. |
+| **RLHF** | Reinforcement Learning from Human Feedback. Pipeline: SFT → Reward Model → PPO. Original alignment method (InstructGPT). |
+| **Scaling Laws** | Empirical finding that LM loss follows a power law: `L(N) ∝ N^(-α)`. Predicts performance from compute budget. |
+| **Self-Attention** | Attention where queries, keys, and values all come from the same sequence. Each token attends to all tokens in the sequence. |
+| **SFT (Supervised Fine-Tuning)** | Fine-tuning on instruction-response pairs. Transforms base models into helpful assistants. |
+| **Softmax** | `softmax(xᵢ) = exp(xᵢ) / Σ exp(xⱼ)`. Converts raw scores (logits) to a probability distribution. |
+| **Temperature** | Scaling factor applied to logits before softmax during generation. Higher = more random, lower = more deterministic. |
+| **Token** | The atomic unit of text for the model. Can be a character, subword, or word depending on the tokenizer. |
+| **Transformer** | Neural network architecture based on self-attention, introduced in "Attention Is All You Need" (2017). Foundation of all modern LLMs. |
+
+---
+
+## Progress Tracker
+
+Use this checklist to track your progress:
+
+### Phase 1: Foundations ☐
+- [ ] Week 1: Linear algebra & calculus videos complete
+- [ ] Week 1: Implemented `matmul`, `softmax`, `cross_entropy` from scratch
+- [ ] Week 2: Watched 3B1B neural networks series
+- [ ] Week 2: Built micrograd (autograd engine)
+- [ ] Week 3: Completed PyTorch 60-min blitz
+- [ ] Week 3: Built bigram + MLP language models (makemore Parts 1–2)
+
+### Phase 2: Transformer Architecture ☐
+- [ ] Week 4: Completed makemore Parts 3–5
+- [ ] Week 4: Can manually backpropagate through a small network
+- [ ] Week 5: Read "Attention Is All You Need" (all of §3)
+- [ ] Week 5: Can draw the full Transformer architecture from memory
+- [ ] Week 6: Watched "Let's Build GPT" and implemented along
+- [ ] Week 6: Trained a working GPT on Shakespeare that generates text
+
+### Phase 3: Language Modeling ☐
+- [ ] Week 7: Implemented BPE from scratch
+- [ ] Week 7: Trained a HuggingFace tokenizer on custom data
+- [ ] Week 8: Read Chinchilla & GPT-3 papers
+- [ ] Week 8: Can calculate FLOPs and training time for a given model size
+- [ ] Week 9: Pretrained a small GPT (10M–50M params)
+- [ ] Week 9: Pushed model to Hugging Face Hub
+
+### Phase 4: HF Ecosystem ☐
+- [ ] Week 10: Loaded and ran 5 different models via `pipeline()`
+- [ ] Week 11: Fine-tuned a text classifier with `Trainer`
+- [ ] Week 11: Model pushed to Hub
+- [ ] Week 12: Deployed a Gradio demo on HF Spaces
+
+### Phase 5: Fine-Tuning & Alignment ☐
+- [ ] Week 13: SFT'd SmolLM2 into a chat model
+- [ ] Week 14: Applied QLoRA to a 1.7B model
+- [ ] Week 15: Trained a DPO-aligned model
+- [ ] Week 16: Trained a GRPO reasoning model
+
+### Phase 6: Advanced ☐
+- [ ] Week 17: Benchmarked all models with `lighteval`
+- [ ] Week 18: Generated synthetic data, quantized a model
+- [ ] Week 19: Built a RAG agent with smolagents
+- [ ] Week 20: Completed capstone project
+
+---
+
+> *"The best way to understand LLMs is to build one from scratch. The second best way is to train one. The third best way is to read the papers. Do all three."*
+
+**Good luck on your journey! 🚀**