LLM

What’s an ML model?

An object/file that has been conditioned by an algorithm to learn specific patterns in datasets

What’s an LLM?

A machine learning model designed to understand and generate human language
Trained on vast amounts of text data

Today’s most powerful open-weights model is Meta’s Llama3.1

LLM’s are made up of two files, parameters and run.c

• parameters contains the weights of the language model. For example, Llama3:8b has a parameter size of 8 billion, so the parameters file will be 16GB (you need two bytes (float16) to store each parameter).

• run.c contains ~500 lines of C code to implement the NN architecture that uses the parameters to run the model

---

How to train an LLM from scratch:

1. Pretraining (base model)

   1. Download ~10TB text
   2. Buy cluster of ~6,000 GPUs
   3. Lossy compress text into a neural network (~$2M, ~12 days)
   4. Obtain base model

2. Finetuning (assistant model)

   1. Write labeling instructions
   2. Hire people to write quality Q&A responses (and/or comparisons (RLHF))
   3. “Finetune” base model on this data (~1 day)
   4. Obtain assistant model
   5. Run evaluations
   6. Deploy
   7. Monitor, collect misbehaviors, go to step 1

---

Todo: expand on the following:

• Closed-source models
• Llama3.1
• OpenWebUI
• Ollama