🗣️ Large Language Model Course

The LLM course is divided into three parts:

1. 🧩 LLM Fundamentals covers essential knowledge about mathematics, Python, and neural networks.
2. 🧑‍🔬 The LLM Scientist focuses on learning how to build the best possible LLMs using the latest techniques
3. 👷 The LLM Engineer focuses on how to create LLM-based solutions and deploy them.

📝 Notebooks

A list of notebooks and articles related to large language models.

Fine-tuning

Quantization

Other

🧩 LLM Fundamentals

1. Mathematics for Machine Learning

Before mastering machine learning, it is important to understand the fundamental mathematical concepts that power these algorithms.

• Linear Algebra: This is crucial for understanding many algorithms, especially those used in deep learning. Key concepts include vectors, matrices, determinants, eigenvalues and eigenvectors, vector spaces, and linear transformations.
• Calculus: Many machine learning algorithms involve the optimization of continuous functions, which requires an understanding of derivatives, integrals, limits, and series. Multivariable calculus and the concept of gradients are also important.
• Probability and Statistics: These are crucial for understanding how models learn from data and make predictions. Key concepts include probability theory, random variables, probability distributions, expectations, variance, covariance, correlation, hypothesis testing, confidence intervals, maximum likelihood estimation, and Bayesian inference.

📚 Resources:

• 3Blue1Brown – The Essence of Linear Algebra: Series of videos that give a geometric intuition to these concepts.
• StatQuest with Josh Starmer – Statistics Fundamentals: Offers simple and clear explanations for many statistical concepts.
• AP Statistics Intuition by Ms Aerin: List of Medium articles that provide the intuition behind every probability distribution.
• Immersive Linear Algebra: Another visual interpretation of linear algebra.
• Khan Academy – Linear Algebra: Great for beginners as it explains the concepts in a very intuitive way.
• Khan Academy – Calculus: An interactive course that covers all the basics of calculus.
• Khan Academy – Probability and Statistics: Delivers the material in an easy-to-understand format.

2. Python for Machine Learning

Python is a powerful and flexible programming language that’s particularly good for machine learning, thanks to its readability, consistency, and robust ecosystem of data science libraries.

• Python Basics: Python programming requires a good understanding of the basic syntax, data types, error handling, and object-oriented programming.
• Data Science Libraries: It includes familiarity with NumPy for numerical operations, Pandas for data manipulation and analysis, Matplotlib and Seaborn for data visualization.
• Data Preprocessing: This involves feature scaling and normalization, handling missing data, outlier detection, categorical data encoding, and splitting data into training, validation, and test sets.
• Machine Learning Libraries: Proficiency with Scikit-learn, a library providing a wide selection of supervised and unsupervised learning algorithms, is vital. Understanding how to implement algorithms like linear regression, logistic regression, decision trees, random forests, k-nearest neighbors (K-NN), and K-means clustering is important. Dimensionality reduction techniques like PCA and t-SNE are also helpful for visualizing high-dimensional data.

📚 Resources:

• Real Python: A comprehensive resource with articles and tutorials for both beginner and advanced Python concepts.
• freeCodeCamp – Learn Python: Long video that provides a full introduction into all of the core concepts in Python.
• Python Data Science Handbook: Free digital book that is a great resource for learning pandas, NumPy, Matplotlib, and Seaborn.
• freeCodeCamp – Machine Learning for Everybody: Practical introduction to different machine learning algorithms for beginners.
• Udacity – Intro to Machine Learning: Free course that covers PCA and several other machine learning concepts.

3. Neural Networks

Neural networks are a fundamental part of many machine learning models, particularly in the realm of deep learning. To utilize them effectively, a comprehensive understanding of their design and mechanics is essential.

• Fundamentals: This includes understanding the structure of a neural network such as layers, weights, biases, activation functions (sigmoid, tanh, ReLU, etc.)
• Training and Optimization: Familiarize yourself with backpropagation and different types of loss functions, like Mean Squared Error (MSE) and Cross-Entropy. Understand various optimization algorithms like Gradient Descent, Stochastic Gradient Descent, RMSprop, and Adam.
• Overfitting: Understand the concept of overfitting (where a model performs well on training data but poorly on unseen data) and learn various regularization techniques (dropout, L1/L2 regularization, early stopping, data augmentation) to prevent it.
• Implement a Multilayer Perceptron (MLP): Build an MLP, also known as a fully connected network, using PyTorch.

📚 Resources:

• 3Blue1Brown – But what is a Neural Network?: This video gives an intuitive explanation of neural networks and their inner workings.
• freeCodeCamp – Deep Learning Crash Course: This video efficiently introduces all the most important concepts in deep learning.
• Fast.ai – Practical Deep Learning: Free course designed for people with coding experience who want to learn about deep learning.
• Patrick Loeber – PyTorch Tutorials: Series of videos for complete beginners to learn about PyTorch.

4. Natural Language Processing (NLP)

NLP is a fascinating branch of artificial intelligence that bridges the gap between human language and machine understanding. From simple text processing to understanding linguistic nuances, NLP plays a crucial role in many applications like translation, sentiment analysis, chatbots, and much more.

• Text Preprocessing: Learn various text preprocessing steps like tokenization (splitting text into words or sentences), stemming (reducing words to their root form), lemmatization (similar to stemming but considers the context), stop word removal, etc.
• Feature Extraction Techniques: Become familiar with techniques to convert text data into a format that can be understood by machine learning algorithms. Key methods include Bag-of-words (BoW), Term Frequency-Inverse Document Frequency (TF-IDF), and n-grams.
• Word Embeddings: Word embeddings are a type of word representation that allows words with similar meanings to have similar representations. Key methods include Word2Vec, GloVe, and FastText.
• Recurrent Neural Networks (RNNs): Understand the working of RNNs, a type of neural network designed to work with sequence data. Explore LSTMs and GRUs, two RNN variants that are capable of learning long-term dependencies.

📚 Resources:

• RealPython – NLP with spaCy in Python: Exhaustive guide about the spaCy library for NLP tasks in Python.
• Kaggle – NLP Guide: A few notebooks and resources for a hands-on explanation of NLP in Python.
• Jay Alammar – The Illustration Word2Vec: A good reference to understand the famous Word2Vec architecture.
• Jake Tae – PyTorch RNN from Scratch: Practical and simple implementation of RNN, LSTM, and GRU models in PyTorch.
• colah’s blog – Understanding LSTM Networks: A more theoretical article about the LSTM network.

🧑‍🔬 The LLM Scientist

1. The LLM architecture

While an in-depth knowledge about the Transformer architecture is not required, it is important to have a good understanding of its inputs (tokens) and outputs (logits). The vanilla attention mechanism is another crucial component to master, as improved versions of it are introduced later on.

• High-level view: Revisit the encoder-decoder Transformer architecture, and more specifically the decoder-only GPT architecture, which is used in every modern LLM.
• Tokenization: Understand how to convert raw text data into a format that the model can understand, which involves splitting the text into tokens (usually words or subwords).
• Attention mechanisms: Grasp the theory behind attention mechanisms, including self-attention and scaled dot-product attention, which allows the model to focus on different parts of the input when producing an output.
• Text generation: Learn about the different ways the model can generate output sequences. Common strategies include greedy decoding, beam search, top-k sampling, and nucleus sampling.

📚 References:

• The Illustrated Transformer by Jay Alammar: A visual and intuitive explanation of the Transformer model.
• The Illustrated GPT-2 by Jay Alammar: Even more important than the previous article, it is focused on the GPT architecture, which is very similar to Llama’s.
• LLM Visualization by Brendan Bycroft: Incredible 3D visualization of what happens inside of an LLM.

• nanoGPT by Andrej Karpathy: A 2h-long YouTube video to reimplement GPT from scratch (for programmers).
• Attention? Attention! by Lilian Weng: Introduce the need for attention in a more formal way.
• Decoding Strategies in LLMs: Provide code and a visual introduction to the different decoding strategies to generate text.

2. Building an instruction dataset

While it’s easy to find raw data from Wikipedia and other websites, it’s difficult to collect pairs of instructions and answers in the wild. Like in traditional machine learning, the quality of the dataset will directly influence the quality of the model, which is why it might be the most important component in the fine-tuning process.

• Alpaca-like dataset: Generate synthetic data from scratch with the OpenAI API (GPT). You can specify seeds and system prompts to create a diverse dataset.
• Advanced techniques: Learn how to improve existing datasets with Evol-Instruct, how to generate high-quality synthetic data like in the Orca and phi-1 papers.
• Filtering data: Traditional techniques involving regex, removing near-duplicates, focusing on answers with a high number of tokens, etc.
• Prompt templates: There’s no true standard way of formatting instructions and answers, which is why it’s important to know about the different chat templates, such as ChatML, Alpaca, etc.

📚 References:

• Preparing a Dataset for Instruction tuning by Thomas Capelle: Exploration of the Alpaca and Alpaca-GPT4 datasets and how to format them.
• Generating a Clinical Instruction Dataset by Solano Todeschini: Tutorial on how to create a synthetic instruction dataset using GPT-4.
• GPT 3.5 for news classification by Kshitiz Sahay: Use GPT 3.5 to create an instruction dataset to fine-tune Llama 2 for news classification.
• Dataset creation for fine-tuning LLM: Notebook that contains a few techniques to filter a dataset and upload the result.
• Chat Template by Matthew Carrigan: Hugging Face’s page about prompt templates

3. Pre-training models

Pre-training is a very long and costly process, which is why this is not the focus of this course. It’s good to have some level of understanding of what happens during pre-training, but hands-on experience is not required.

• Data pipeline: Pre-training requires huge datasets (e.g., Llama 2 was trained on 2 trillion tokens) that need to be filtered, tokenized, and collated with a pre-defined vocabulary.
• Causal language modeling: Learn the difference between causal and masked language modeling, as well as the loss function used in this case. For efficient pre-training, learn more about Megatron-LM or gpt-neox.
• Scaling laws: The scaling laws describe the expected model performance based on the model size, dataset size, and the amount of compute used for training.
• High-Performance Computing: Out of scope here, but more knowledge about HPC is fundamental if you’re planning to create your own LLM from scratch (hardware, distributed workload, etc.).

📚 References:

• LLMDataHub by Junhao Zhao: Curated list of datasets for pre-training, fine-tuning, and RLHF.
• Training a causal language model from scratch by Hugging Face: Pre-train a GPT-2 model from scratch using the transformers library.
• TinyLlama by Zhang et al.: Check this project to get a good understanding of how a Llama model is trained from scratch.
• Causal language modeling by Hugging Face: Explain the difference between causal and masked language modeling and how to quickly fine-tune a DistilGPT-2 model.
• Chinchilla’s wild implications by nostalgebraist: Discuss the scaling laws and explain what they mean to LLMs in general.
• BLOOM by BigScience: Notion pages that describes how the BLOOM model was built, with a lot of useful information about the engineering part and the problems that were encountered.
• OPT-175 Logbook by Meta: Research logs showing what went wrong and what went right. Useful if you’re planning to pre-train a very large language model (in this case, 175B parameters).
• LLM 360: A framework for open-source LLMs with training and data preparation code, data, metrics, and models.

4. Supervised Fine-Tuning

Pre-trained models are only trained on a next-token prediction task, which is why they’re not helpful assistants. SFT allows you to tweak them into responding to instructions. Moreover, it allows you to fine-tune your model on any data (private, not seen by GPT-4, etc.) and use it without having to pay for an API like OpenAI’s.

• Full fine-tuning: Full fine-tuning refers to training all the parameters in the model. It is not an efficient technique, but it produces slightly better results.
• LoRA: A parameter-efficient technique (PEFT) based on low-rank adapters. Instead of training all the parameters, we only train these adapters.
• QLoRA: Another PEFT based on LoRA, which also quantizes the weights of the model in 4 bits and introduce paged optimizers to manage memory spikes. Combine it with Unsloth to run it efficiently on a free Colab notebook.
• Axolotl: A user-friendly and powerful fine-tuning tool that is used in a lot of state-of-the-art open-source models.
• DeepSpeed: Efficient pre-training and fine-tuning of LLMs for multi-GPU and multi-node settings (implemented in Axolotl).

📚 References:

• The Novice’s LLM Training Guide by Alpin: Overview of the main concepts and parameters to consider when fine-tuning LLMs.
• LoRA insights by Sebastian Raschka: Practical insights about LoRA and how to select the best parameters.
• Fine-Tune Your Own Llama 2 Model: Hands-on tutorial on how to fine-tune a Llama 2 model using Hugging Face libraries.
• Padding Large Language Models by Benjamin Marie: Best practices to pad training examples for causal LLMs
• A Beginner’s Guide to LLM Fine-Tuning: Tutorial on how to fine-tune a CodeLlama model using Axolotl.

5. Reinforcement Learning from Human Feedback

After supervised fine-tuning, RLHF is a step used to align the LLM’s answers with human expectations. The idea is to learn preferences from human (or artificial) feedback, which can be used to reduce biases, censor models, or make them act in a more useful way. It is more complex than SFT and often seen as optional.

• Preference datasets: These datasets typically contain several answers with some kind of ranking, which makes them more difficult to produce than instruction datasets.
• Proximal Policy Optimization: This algorithm leverages a reward model that predicts whether a given text is highly ranked by humans. This prediction is then used to optimize the SFT model with a penalty based on KL divergence.
• Direct Preference Optimization: DPO simplifies the process by reframing it as a classification problem. It uses a reference model instead of a reward model (no training needed) and only requires one hyperparameter, making it more stable and efficient.

📚 References:

• An Introduction to Training LLMs using RLHF by Ayush Thakur: Explain why RLHF is desirable to reduce bias and increase performance in LLMs.
• Illustration RLHF by Hugging Face: Introduction to RLHF with reward model training and fine-tuning with reinforcement learning.
• StackLLaMA by Hugging Face: Tutorial to efficiently align a LLaMA model with RLHF using the transformers library.
• LLM Training: RLHF and Its Alternatives by Sebastian Rashcka: Overview of the RLHF process and alternatives like RLAIF.
• Fine-tune Mistral-7b with DPO: Tutorial to fine-tune a Mistral-7b model with DPO and reproduce NeuralHermes-2.5.

6. Evaluation

Evaluating LLMs is an undervalued part of the pipeline, which is time-consuming and moderately reliable. Your downstream task should dictate what you want to evaluate, but always remember the Goodhart’s law: “when a measure becomes a target, it ceases to be a good measure.”

• Traditional metrics: Metrics like perplexity and BLEU score are not popular as they were because they’re flawed in most contexts. It is still important to understand them and when they can be applied.
• General benchmarks: Based on the Language Model Evaluation Harness, the Open LLM Leaderboard is the main benchmark for general-purpose LLMs (like ChatGPT). There are other popular benchmarks like BigBench, MT-Bench, etc.
• Task-specific benchmarks: Tasks like summarization, translation, question answering have dedicated benchmarks, metrics, and even subdomains (medical, financial, etc.), such as PubMedQA for biomedical question answering.
• Human evaluation: The most reliable evaluation is the acceptance rate by users or comparisons made by humans. If you want to know if a model performs well, the simplest but surest way is to use it yourself.

📚 References:

• Perplexity of fixed-length models by Hugging Face: Overview of perplexity with code to implement it with the transformers library.
• BLEU at your own risk by Rachael Tatman: Overview of the BLEU score and its many issues with examples.
• A Survey on Evaluation of LLMs by Chang et al.: Comprehensive paper about what to evaluate, where to evaluate, and how to evaluate.
• Chatbot Arena Leaderboard by lmsys: Elo rating of general-purpose LLMs, based on comparisons made by humans.
• LLM AutoEval: Automatically evaluate LLMs from a Colab notebook.

7. Quantization

Quantization is the process of converting the weights (and activations) of a model using a lower precision. For example, weights stored using 16 bits can be converted into a 4-bit representation. This technique has become increasingly important to reduce the computational and memory costs associated to LLMs.

• Base techniques: Learn the different levels of precision (FP32, FP16, INT8, etc.) and how to perform naïve quantization with absmax and zero-point techniques.
• GGUF and llama.cpp: Originally designed to run on CPUs, llama.cpp and the GGUF format have become the most popular tools to run LLMs on consumer-grade hardware.
• GPTQ and EXL2: GPTQ and, more specifically, the EXL2 format offer an incredible speed but can only run on GPUs. Models also take a long time to be quantized.
• AWQ: This new format is more accurate than GPTQ (lower perplexity) but uses a lot more VRAM and is not necessarily faster.

📚 References:

• Introduction to quantization: Overview of quantization, absmax and zero-point quantization, and LLM.int8() with code.
• Quantize Llama models with llama.cpp: Tutorial on how to quantize a Llama 2 model using llama.cpp and the GGUF format.
• 4-bit LLM Quantization with GPTQ: Tutorial on how to quantize an LLM using the GPTQ algorithm with AutoGPTQ.
• ExLlamaV2: The Fastest Library to Run LLMs: Guide on how to quantize a Mistral model using the EXL2 format and run it with the ExLlamaV2 library.
• Understanding Activation-Aware Weight Quantization by FriendliAI: Overview of the AWQ technique and its benefits.

8. Inference optimization

Numerous optimization techniques have been developed to reduce VRAM usage and enhance the speed of generation. In addition to quantization methods, these improvements frequently involve the implementation of more efficient attention mechanisms and architectural changes.

• Flash Attention: Optimization of the attention mechanism to transform its complexity from quadratic to linear, speeding up both training and inference.
• Key-value cache: Understand the key-value cache and the improvements introduced in Multi-Query Attention (MQA) and Grouped-Query Attention (GQA).
• Speculative decoding: Use a small model to produce drafts that are then reviewed by a larger model to speed up text generation.
• Positional encoding: Understand positional encodings in transformers, particularly relative schemes like RoPE, ALiBi, and YaRN. (Not directly connected to inference optimization but to longer context windows.)

📚 References:

• GPU Inference by Hugging Face: Explain how to optimize inference on GPUs.
• Optimizing LLMs for Speed and Memory by Hugging Face: Explain three main techniques to optimize speed and memory, namely quantization, Flash Attention, and architectural innovations.
• Assisted Generation by Hugging Face: HF’s version of speculative decoding, it’s an interesting blog post about how it works with code to implement it.
• Extending the RoPE by EleutherAI: Article that summarizes the different position-encoding techniques.
• Extending Context is Hard… but not Impossible by kaiokendev: This blog post introduces the SuperHOT technique and provides an excellent survey of related work.

Acknowledgements

This work is derived from mlabonne in Github.

This roadmap was inspired by the excellent DevOps Roadmap from Milan Milanović and Romano Roth.

Special thanks to Thomas Thelen for motivating me to create a roadmap, and André Frade for his input and review of the first draft.

Disclaimer: I am not affiliated with any sources listed here.