Stop 10 · NLP & Modern AITopics · 6/7

Pretrained LLMs & Fine-Tuning Basics

In short

Training a large language model from scratch costs a fortune. Almost everyone starts from a pretrained model and adapts it with prompts or a little fine-tuning.

Using an already-trained large language model (via an API or open weights) and adapting it to a specific task — often via fine-tuning or, more commonly today, careful prompting.

The mascot in a graduation cap receives a violet binder from an older mentor at a bank counter.

Think of it like this

The fresh graduate's induction

A fresh graduate joins a bank. They already read, write and reason well from years of study. The bank doesn't reteach English. It runs a two-week induction on its own forms and rules. Pretraining is the degree. Fine-tuning is the induction.

Watch it happen

From general model to your task

Pretrained

Try prompts

Label data

Fine-tune

Evaluate

1/5A company trained the model on huge amounts of text to predict the next token. That's pretraining.

The big ideas

What you need to know

Pretraining is general

The model learns grammar, facts and patterns from a vast text collection. It's expensive and done once.

Fine-tuning is narrow

A smaller dataset teaches a style, format or label set, such as classifying support tickets.

LoRA keeps it cheap

Instead of updating billions of weights, LoRA trains small extra matrices. It fits on much smaller GPUs.

Facts go stale

Fine-tuning is poor at adding fresh knowledge. For changing facts, retrieval from your documents usually works better.

See it in code

Walk through the code

python

1from datasets import load_dataset2from transformers import (AutoModelForSequenceClassification, AutoTokenizer,3                          Trainer, TrainingArguments)4 5name = "distilbert-base-uncased"6tok = AutoTokenizer.from_pretrained(name)7model = AutoModelForSequenceClassification.from_pretrained(name, num_labels=2)8 9data = load_dataset("imdb")10data = data.map(lambda b: tok(b["text"], truncation=True), batched=True)11 12args = TrainingArguments("out", num_train_epochs=1, per_device_train_batch_size=16)13trainer = Trainer(model=model, args=args, processing_class=tok,14                  train_dataset=data["train"].shuffle(seed=0).select(range(2000)),15                  eval_dataset=data["test"].shuffle(seed=0).select(range(500)))16trainer.train()

1/4Load a small pretrained model and add a new 2-label classification head.

Fine-tuning a small pretrained model for movie-review sentiment with Hugging Face. Run it on a free Colab GPU.

Save yourself the pain

Common mistakes

Oops

Jumping to fine-tuning before trying good prompts.

Do this instead

Prompt first and measure. Fine-tune only when prompts can't hit the quality, cost or speed you need.

Oops

Fine-tuning on a few dozen messy examples and expecting magic.

Do this instead

Clean, consistent, varied examples matter more than quantity. Hold some out to test.

Your mission

Zero-shot vs fine-tuned

~45 min

Quick check

Your support bot needs this week's new policy details. What usually works better than fine-tuning?

Go deeper

1Use a pretrained model (via Hugging Face or an API) for a simple text classification task.
2Learn the difference between full fine-tuning, parameter-efficient fine-tuning, and prompting.
3Understand why fine-tuning from scratch is rarely necessary for most practical tasks today.