Glossary - LLMs Unplugged Facilitator Guide

Role Glossary¶

helper: people who assist during a workshop or training, putting up stickies, passing out materials, secondary training during parallel. helpers at a training may be new facilitators or even trainees
facilitator: person or people (co-facilitators) who are running a workshop. should have some more technical depth in understanding AI and its risks and completed this training
participant: person who attends the workshop to learn what AI is.
trainee: person who attends the training to become a facilitator in the future
trainer: person who runs the training; an experienced facilitator with deeper background to help prepare people to run the workshop on their own.
training: event to prepare future facilitators
workshop: the LLM unplugged workshop, with a targt audience of no tech background required

Technical Terms¶

agent: an AI that is able to do multiple things (FIXME-anthropic defs)
algorithm: a set of instructions to complete a task; a recipe
anonymous function: a function that’s defined on the fly, typically to lighten syntax or return a function within a function. In python, they’re defined with the lambda keyword.; docs
artificial intelligence: the study of intelligent agents that receive precepts from the environment and take action; a technology that has input ouptut behavior that appears as an intelligent agent; a sub field of computer science
artificial neuron: a type of function that takes a weighted sum of the inputs and passes them to an activation function; a type of mathematical, or computer, function that approximates some functions of biological neurons
attention layer: a type of neural network layer that assigned weights to the token embedding to represent the context of the current token
binary: a place based number system with base 2
categorical variable: a variable that can only take one of a fixed discrete number of values, in general unordered
classification: a type of machine learning where a categorical target variable is predicted from features
class: a value of the target variable
cluster: a group of samples that is similar by some definition
clustering: a type of unsupervised learning that finds groups or clusters among the samples compute; apply a series of mathematical operations to an input
deep learning: a type of machine learning that consists of creating large numbers of hidden layers of artificial neurons
development: a stage in the “life cycle” of code wheere is it not necessarily ready for users, but the developers are actively working on it.
discriminative model: a type of statistical model that describes the decision boundary to make a prediction
embedding: a numerical representation for complex data, for example text or images
feature: in machine learning, one input to the prediction algorithm, typically of many, stored together in a feature vector
generative model: a type of statistical model that describes the way that the underlying data is distributed with sufficient detail to be able to sample new data that is in the same distribution as the original data; often in
generative pretrained transformer: It is a transformer model that takes in tokens and can be sampled from to to produce novel token sequences that are similar to its training data; a specific type of generative model
label: the output value for a particular sample or of a prediction algorithm (ie true label or predicted label)
lambda: the keyword used to define an anonymous function; lambda functions are defined with a compact syntax <name> = lambda <parameters>: <body>; docs
language: a vocabulary and a grammar for how to combine the words in order to convey meaning
large language model: a probability distribution that describes the structure of sequences of tokens and enables next token prediction as well as serves as a mathematical description of tokens using many parameters
learning algorithm: an algorithm for finding patterns in data; a paramter optimization or parameter fitting algorithm
machine learning: a sub field of AI; emphasizes the use of data to learn relationships, studies the algorithms that find patterns in data
model: a simplification of some part of the world
model class: a generic mathematical model that can be fit to a specific part of the world by setting parameter values
neural network: a set of artificial neurons where the outputs of some are the inputs to the others, serving to “connect” them as a set, or network; often described as layers where connections are from one layer to the next.
neural network architecture: the general structure of how a neural network is set up, not the specific weights, but the size of the layers, number of layers, types of layers for example
number: a mathematical object used to count, measure and label
number system: a cultural artifact that describes the symbols used to represent quantities and how to interpret them as numbers
parameter: (general programming) all inputs to a function; (in ML) the values that transform a generic function into a specific function
place based: a type of number system where the position or place of the symbol changes its meaning. So in 1, 10, and 100, the symbol 1 represents different values because of its position.
predictive model: a mathemtical model, typically implemented in computer code, that predicts a label for an input or an outcome based on past
probability distribution: a function that takes an event (or value) as input and returns the likelihood of that even occuring
production: a state of a system where it is in use by users unaffiliated with the developers and generally unrestricted (except for possibly payment); contrast with research, sandbox, and development
programming language: a human readable language that can be processed by computers and translated (either live through an interpreter or in batch through a compiler) to machine instructions that the physical computer can execute
machine representation: a description of an object that can be stored in a computer
regression: supervised learning that predicts a continuous valued output
research: a state of a system where is has been built and evaluated, but not deployed to general users, instead is only available to the team that built it and (potentially) other authorized users; contrast with production
retrieval agumented generation: a process to produce text that is grounded in content from specific resources in order to improve the factual correctness of the text or augment with specific data that was not available to a base model at the time of traning; commonly used to “chat with” a document or to prevent false statements
sandbox: a state of a system where is is available to test, but is not officially deployed, may have simulated data instead of real data or be isolated from external systems
statistical model: a type of mathematical model that describes its object of inquiry using probability distributions
supervised learning: a type of machine learning that requires both features and target variables at time of training; machine learning with labeled examples
target: the variable to be predicted in a machine learning problem setup, the dependent variable; the output of a prediction algorithm; also called the dependent variable or label
test accuracy: percentage of predictions that the model predict correctly, based on held-out (previously unseen) test data
test data: data that was not used in training that is instead used to evaluate the perforance of a model
token: the basic units used in computational processing of texts from human spoken (not programming) languages; a part of a word or a word, approximately like prefix, base, suffix parts, in general 1000 tokens is about 750 words
transformer: a type of model, specifically, a type of neural network architecture
training data: data used to find the values of the model parameters to make the model work.
unsupervised learning: a type of machine learning that does not use target variables at learning (fit) time.; machine learning from unlabeled examples