AI Training Curriculum

Created August 2023
Updated May 2024

    Do you want to learn about LLM (Large Langiuage Models)?
    AI ChatBots ?
    RAG systems (Retrieval Augmented Geenration) ?

Here is a short tutorial.

Prerequisites:

Computer (desktop or laptop)
with Unix OS - (because all AI is running on Unix)

You have 3 choices:

- Linux standalone (or dual-boot) - Ubuntu, Mint, ...
- Linux under Windows (WSL2)
- Mac with Apple Silicone chip (M1, M2, M3, ...)

For Linux you need a CUDA-compatible Nvidia GPU.
with as much memory as you can afford.

Nvidia RTX 3090 or 4090 with 24GB memory ($800 - $1,600)

Search for instructions to activate WSL2 (Windows Subsystem for Linux version 2)

• https://learn.microsoft.com/en-us/windows/wsl/install
• https://learn.microsoft.com/en-us/windows/wsl/tutorials/wsl-vscode

For Mac you should get a macbook with M1, M2, or M3 chip.
At least 16GB memory (but 64 or 96 or 128GB is better).
And at least 1TB SSD

Here are some affordable options for you (on ebay, May 2024):

• $900 - MacBook Air 13.6" screen, M1 chip with 16GB memory, 1TB SSD
• $1,500 - MacBook Air 15" screen, M2 chip with 16GB memory, 1TB SSD
• $1,600 - MacBook Air 15" screen, M2 chip with 24GB memory, 1TB SSD

Prerequisites (not really required):

• Linux/Unix
• Python (Numpy, Pandas), JSON
• Basic math and statistics
• Basic Machine Learning (regression, classification)
• SQL
• Basic cloud usage: files (AWS s3, Azure data lake, etc),
  SQL databases, ETL tools, Analytics dashboards

As you go through the training materials below, you will have questions.
I recommend you to use ChatBots to find answers:

• ChatGPT ( https://chat.openai.com )
• Gemini Advanced ( https://gemini.google.com )
• Claude ( https://claude.ai )
• Groq platform ( https://groq.com )
• Huggingface ( https://huggingface.co/chat )
• Google search

Lesson 1 - Introduction into Generative AI

Most popular model types - Transformer, GAN, Diffusion.

Below are links to some materials (text, videos)
It should take you 2-3 hours max
Then we can discuss on the call

Start with watching this my old video from June 9, 2023:
"Generative AI - Transformers, GANs, Stable Diffusion."

• https://www.youtube.com/watch?v=CA0Yfds-nlc

Also download updated PPT for this lecture here:

• https://github.com/lselector/seminar/tree/master/2023

Then read this illustrated description (and watch video) by Jay Alammar:

• http://jalammar.github.io/illustrated-transformer/
• https://www.youtube.com/watch?v=-QH8fRhqFHM

Excellent animated tutorials about deep learning.
You may start with #5 and #6 explaining GPT and Transformers.

• https://www.youtube.com/watch?v=aircAruvnKk&list=PLZHQObOWTQDNU6R1_67000Dx_ZCJB-3pi

How ChatGPT Works Technically | ChatGPT Architecture (7min):

• https://www.youtube.com/watch?v=bSvTVREwSNw

Look at leaderboards :

• https://chat.lmsys.org/?leaderboard
• https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard

Some terms to know:

    - LLM = "Large Language Model"
    - word2vec (2013) 
    - Google Translate Paper "Attention is all you need", 
      attention mechanism
      transformer (Google 2016-2017)
    - tokenizer: split text into tokens
      tokens are ~ 4 chars.
      dictionary of tokens may have ~ 30K .. 120K entries
    - embeddings - numeric vectors used to represent tokens 
      or words or even chunks of text
    - encoder and decoder layers 
    - softmax function
    - Google Translate (encoder+decoder) 
      vs BERT (encoder only) 
      vs ChatGPT (decoder only)

Lesson 2 - Recent AI updates

Watch some recent AI updates on my youtube channel:

• https://www.youtube.com/@lev-selector

You can download all my slides here as one zip file:

• https://github.com/lselector/seminar

I recommend the following lectures:

AI Training (Brief Introduction) - June 21, 2023

• https://www.youtube.com/watch?v=W_2Vb9aBkao

Andrej Karpathy - Intro into LLMs - Nov 22, 2023

• https://www.youtube.com/watch?v=zjkBMFhNj_g

AI Updates January 19, 2024

• https://www.youtube.com/watch?v=0R5glMg69I4&t=183s

RAG = Retrieval Augmented Generation

1. convert your text into vectors in vector database 
2. convert your question into vector 
3. do vector similarity search - retrieve best matches 
4. re-rank/sort the findings 
5. use LLM to convert the findings into response

Advance RAG

• https://luv-bansal.medium.com/advance-rag-improve-rag-performance-208ffad5bb6a

Mastering RAG

• https://www.rungalileo.io/blog/mastering-rag-how-to-architect-an-enterprise-rag-system

RAG on AWS:

• https://aws.amazon.com/blogs/machine-learning/quickly-build-high-accuracy-generative-ai-applications-on-enterprise-data-using-amazon-kendra-langchain-and-large-language-models/

Lesson 3 - Ollama, LM Studio - running LLM locally

Ollama

Ollama	In Terminal

Download and install Ollama from https://ollama.ai

Open terminal window (or CMD window on Windows) and type:

ollama run llama3:latest

use terminal to chat to local model press /bye to exit

LM Studio

Download and install LM Studio from https://lmstudio.ai

Use it to download and run LLMs locally

Lesson 4 - LangChain

Python framework to work with LLMs

Langchain was first released in October 2022
It was created by Harrison Chase
Written in Python and JavaScript

• https://langchain.com
• https://github.com/langchain-ai/langchain
• https://en.wikipedia.org/wiki/LangChain
• https://docs.langchain.com
• https://python.langchain.com/docs/get_started
• https://python.langchain.com/docs/modules/model_io/prompts/prompt_templates/
• https://github.com/MarekOo/ChatGPTLangChain

Andrew Ng short Coursera courses:
https://learn.deeplearning.ai

• ChatGPT Prompt Engineering for Developers
• LangChain for LLM Application Development
• How Diffusion Models Work
• Building Systems with the ChatGPT API
• LangChain Chat with Your Data
• Building Generative AI Applications with Gradio ...

Search YouTube for "langchain tutorial", for example:

Using ChatGPT with YOUR OWN Data. This is magical. (LangChain OpenAI API)
https://www.youtube.com/watch?v=9AXP7tCI9PI

LangChain Explained in 13 Minutes | QuickStart Tutorial for Beginners
https://www.youtube.com/watch?v=aywZrzNaKjs

LangChain provides:

• Universal API for LLMs (GPT-3, BLOOM, and Jurassic-1 Jumbo, ...)
• Chains = sequences of commands for LLM
• End-to-end chains for popular apps (chatbots, question-answering, and summarization)
• Memory (keeping info about previous chat messages, ...)
• Tools for debugging, testing, evaluating, and monitoring LLM apps
• Prompt templates - strings containing variables in curly braces {myvar}. For example, templates for:
  • chatbots
  • ELI5 question-answering ("Explain Like I'm Five")
  • summarization
  • etc
• Agents - use LLMs to decide what actions should be taken (generate a plan or execute tasks)

Lesson 5 - Simple Chatbots

Review several simple exampels of ChatBots
under subdirectory "mychat" here:

app_chainlit.py
app_chainlit_ollama.py
app_flask.py*
app_streamlit.py
app_streamlit_cookies.py
chainlit.md
images/
templates/

Also please watch this seminar from July 14 where Malte shows chatting using LangChain, OpenAI, and ChromaDB:

• https://www.youtube.com/watch?v=8h5WBHqfoA8
  slides and notebook for this presentation are here:
• https://github.com/lselector/seminar/tree/master/2023/2023-07-14-Embeddings-Malte

Coding tasks:

• reproduce simple Chatbots under "mychat" directory
• reproduce code from one of the tutorials using ChatGPT API and LangChain

Another example:
https://medium.com/@onkarmishra/using-langchain-for-question-answering-on-own-data-3af0a82789ed

Running model on Windows:
https://medium.com/@sasika.roledene/unlocking-llm-running-llama-2-70b-on-a-gpu-with-langchain-561adc616b16

Lesson 6 - Switch from Chroma DB to PostreSQL

We want to switch from Crhoma (temporary in-memory vector db)
to a real local Vector Database - PostreSQL
(LangChain + LLM (ChatGPT) + OpenAI_embedding + PostreSQL)

Tasks:

install PostgreSQL locally

install pgvector extension (slower, but can use L2 distance, inner product, and cosine distance)

install pg_embedding for fast queries (L2 distance only)

• https://neon.tech/blog/pg-embedding-extension-for-vector-search
• https://python.langchain.com/docs/modules/data_connection/text_embedding/
• https://python.langchain.com/docs/integrations/vectorstores/pgembedding

learn to use them, for example:

SELECT * FROM my_table 
WHERE vector @> pg_embedding_search(vector, 0.5, 10);

SELECT * FROM my_table 
WHERE vector @> pgvector_search(vector, 'l2', 0.5, 10);

reproduce the examples from previous lesson using langchain to

• split the document into chunks,
• convert them to embeddings (vectors)
• save these vectors into PostreSQL Database
• use ChatGPT API to answer questions using embeddings

Lesson - 7 Change to use local embedding model

Hugging Face MTEB leaderboard MTEB ( Massive Text Embedding Benchmark )
https://huggingface.co/spaces/mteb/leaderboard

There are ~ 100 models.

512 tokens is fine (approx 1 page of text)

Watch this video: "Deploy the No 1 embedding model on Huggingface with python"
https://www.youtube.com/watch?v=ZB1nn3JWyec

Select the model somewhere from the top which is free and can be installed locally
https://python.langchain.com/docs/integrations/text_embedding/

Lesson 8 -Change to use local LLM

So here everything is local and can work without internet:
LangChain + LLM + embedding + PostreSQL

download a LLM from HuggingFace

• https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard

install and run the model locally - test the performance and accuracy

make a copy of the script from Lesson 3 - and change it to use local LLM instead of ChatGPT

Lesson 9 - talk-to-API

Create a new notebook to make a conversational interface to an API.
Using OpenAI:

• https://python.langchain.com/docs/use_cases/apis
• https://www.linkedin.com/pulse/unlocking-power-ai-transforming-your-api-natural-interface-blum/

Using Classifier

• https://community.openai.com/t/how-to-convert-user-input-into-an-action-like-calling-api/178739

Have fun! Ask questions!

Lesson 10 - Make Lang-to-API transformer for financial data reporting and dashboard (charts)

Create a new jupyter notebook where you can print description of a report or chart (or a dashboard of several charts), and the code should execute on this request and produce the report or dashboard right there in the jupyter notebook.

Lesson 11 - Fine-tuning a model Making model smaller

Lesson 12 - Training our own model