Many of us use LLMs but we dont really understand why it works , by way of fine-tuning a LLM I want to firstly inform people on how an LLM actually functions under the hood, why do we even fine-tune them, how is it even comprehending information? and how can we do this for our own purposes using unsloth, an apache 2.0 licensed platform for fine-tuning open source models. this is an outline of what i will cover:

What is an LLM?

• the mathematical magic behind LLMs

• why do we do things how we do them, how is it different from the past?

• "Attention is All You Need" research papers impact on LLMs .

• Attention / memory

Tokenizers

• Subword

• Word

• Character

Why do we fine-tune?

• Instruction tuning

• Bias tuning / remove racism

• Question answering

• Domain specialization

• Reduced cost

What is RAG and how is this different?

• Fine-tuning = update weights

• RAG = external retrieval

Prerequisites for fine-tuning

• A base model (open source and free model to fit in with FOSS's vision)

• Dataset

• Compute

• Framework (Unsloth / PEFT)

Use cases

• Specialize in a domain

• Model distillation

• Reduced cost

• Decensoring (harmful)

Fine-tuning our own model to generate a new episode of Dragon Ball Z

• Dataset → DBZ scripts/fan prompts collection

• Style/format prompts

Evaluation

• Cross-entropy / Perplexity

• LLM/Human as a Judge

• Human eval (pairwise win-rate)

Options for fine-tuning

• Full fine-tuning

• LoRA

• QLoRA

• Unsloth

Where can you go from here?

• Quantization (4-bit, 8-bit)

• Distillation

• Optimized serving

• Fine-tune + RAG hybrids

• Open datasets / community fine-tunes

• how LLMs work

• what is fine-tuning

• fine-tuning our own LLM to generate anime like episodes