💬LLMs
Large language models, end to end · 67 terms
LLM Basics
10What a large language model is and the tokens it reads. Also the limits on how much it can take in.
A very big model trained on huge amounts of text. It writes by guessing the next word.
A big model trained once on lots of data. Then it is reused for many jobs.
A small piece of text, often part of a word, that a model reads.
The tool that cuts text into tokens. It gives each one a number.
Guessing the next token from the text so far. Repeat it and the model writes.
The most text, in tokens, a model can read at one time.
Summing up a long chat so it fits in the window again.
Tags on chat messages. They say who wrote each one: the app, the person or the model.
How many weights a model has, written like 8B for 8 billion.
One fixed version of a hosted AI model, called by its own ID.
Model Types
8The kinds of language model, sorted by how they were trained, how big they are and who can run them.
A model fresh from its first training. It carries text on but does not follow orders.
A base model trained more, so it can follow orders and chat.
A model that thinks step by step before it answers. Slower, but better at hard problems.
A model small enough to run on a laptop or a phone.
A model whose weights you can download and run yourself.
One of the best AI models there is right now. Only big labs make them.
An AI model you can only use online. The maker keeps the model itself.
A model that can take in text, photos and sound, not just one kind.
Transformer Architecture
10The design inside almost every big language model. Also the parts it is built from.
The design inside most big chat models. It is built on attention.
How a model picks which other words matter for each word.
Each word in a text looks at the other words in that same text.
Many attention steps run side by side. Each finds its own links.
A faster way to work out attention on a GPU. The answer stays exact.
The part of a transformer that reads text and finds its meaning.
The part of a transformer that writes. It adds one token at a time.
It adds each token's place in the text, so word order is not lost.
A model split into parts called experts. Each token uses only a few.
The part that picks which experts each token goes to.
Language Model Training
18The rounds of training that turn raw text into a model that follows requests.
The first, huge round of training on lots of text. It makes a base model.
The last stage of training. It turns a model that has read a lot into a helper.
More training on a model that is done, so it gets good at one job.
More training on asks paired with good answers. Then the model does what you ask.
A way to tune a model. Only a small add-on learns, and the rest stays the same.
A cheap way to fine-tune. The big model is locked and small parts are trained.
LoRA on a squeezed 4-bit copy of the model. Now it fits on one card.
A model learns a new task and loses skills it had.
A model learns which answers people like best. Then it gives more of those.
Points for a model each time a check shows it got the right answer.
A model learns from pairs of answers. People picked the best one in each pair.
A model that scores answers the way people would. It guides training.
A small model is trained to copy a big one. It is fast and still good.
Rules that say how much better a model gets as it grows.
Train on about 20 words of text for each weight.
Skills that seem to pop up all at once when a model gets big.
Data made by a computer to train a model. It is not taken from real life.
Models trained on other models' text get worse each round.
Text Generation
16The settings that decide which word a model picks next, how long it writes and how the text reaches you.
How a model picks the next token from its list of chances.
Always picking the one most likely next word. Fast, but dull.
Keeping a few good word paths at once, then taking the best one.
A setting for how random the model's word picks are. Low is steady, high is wild.
Picking only from the top tokens whose chances add up to a set share, like 0.9.
Picking from only the top few tokens, like the best 40.
The same prompt can give a new answer each time you ask.
The model's score for each word it could pick next. Near zero means sure.
The most tokens a model may write in one reply. It stops at the cap.
Text that makes the model stop writing as soon as it appears.
The reply is sent bit by bit as it is written, not all at the end.
How long you wait before the first word of an AI answer shows up.
Saved work for the tokens so far, so each new token is quick.
A small model guesses words ahead, and the big model checks them.
A setting that lets a model think in steps first. Then it answers.
The work a model does while it answers. More of it can mean better answers.
Language Model Limits
5The ways a language model goes wrong even when it sounds sure.
When an AI states something false as if it were true.
The date a model's training text ends. It knows little after that.
When an AI agrees with you or praises you, even when you are wrong.
A model gets worse at its task as the text you give it grows longer.
Models miss facts in the middle of a long text more often.