Glossary

The list below includes short definitions of the terms used across the site. They’re deliberately brief, since each one links to the page that explains the term in context, with examples. Type in the box to filter the list.

Term	Definition
Accuracy	The fraction of all predictions that were right. It can be misleading when one class is much more common than the others. See Evaluation Metrics.
Activation function	A simple function that each neuron applies to the sum of its inputs before passing it on. It is what lets a neural network represent complex patterns. See Basic Concepts.
Adversarial prompting	Writing input designed to make a model behave in ways its designers did not intend. See Hack Merlin.
Annotation	Marking up a text by hand, such as recording every named entity in a letter and its type. Annotations made by people are the usual ground truth for evaluating a model. See Ground-Truth Annotation.
Artificial Intelligence (AI)	The broadest term: a field that aims to build systems that perform tasks we would call intelligent if a person did them. It covers everything from spam filters to language generation, so on its own it says little about how a system works. See Basic Concepts.
Attention	The mechanism in a transformer that weighs every token against every other token at once, so the model can relate words however far apart they are. See Language Models.
Benchmark	A standard test that AI companies report scores on. Benchmarks become targets, and a high score says little about how a model will do on your own material. See LLM Behaviour.
Bias (in training data)	Systematic skew in what a model learns, which comes from what is, and isn’t, in its data, and from choices made in designing it. A model reproduces the patterns of its data, including its gaps and prejudices. See Model Training.
Bias–variance tradeoff	As a model gets more complex, errors from oversimplifying (bias) fall and errors from memorizing (variance) rise, so the lowest total error sits in between. This sense of bias is statistical, and differs from bias in training data. See Model Training.
Black box	A system whose workings you can’t inspect: you can measure what a trained model does, but not read what it learned, because its knowledge is spread across its weights. See Model Training.
Chat assistant	An application, such as ChatGPT, Claude, or Gemini, that lets you hold a conversation with an LLM, and adds its own instructions and tools around it. See Basic Concepts.
Coding agent	An AI tool that works inside your project folder, and can read your files, write and run code, read the errors, and try again until a task is done. See Coding with LLMs.
Cohen’s kappa	A measure of agreement between two annotators that corrects for agreement expected by chance. A kappa of 1 means perfect agreement, and 0 means no more than chance. See Ground-Truth Annotation.
Confidence score	A number from 0 to 1 that a model gives each prediction. Keeping only predictions above a threshold trades recall for precision. See Evaluating and Comparing Models.
Confusion matrix	A table of a model’s results, with what each item actually was in the rows, and what the model said it was in the columns. See Evaluation Metrics.
Context window	The limit on how much text a model can work with at one time. Everything in the model’s input, including the chat history and any documents, has to fit into it. See Prompt Processing.
Deep Learning (DL)	Machine learning with neural networks that have many layers (hence deep). It is behind most recent advances in image recognition, speech, translation, and language generation. See Basic Concepts.
Development set	Data held back from training, and used to check the model’s progress and guide tuning. Also called a validation set. See Model Training.
Embedding	A list of numbers that represents a token, placing it as a point in a space where tokens with related meanings sit close together. See Language Models.
Encoder and decoder	The two halves of a sequence-to-sequence model. The encoder reads the input and turns it into an internal representation of its meaning, and the decoder produces the output from it, one token at a time. See Language Models.
Entity resolution	Working out which different names (surface forms) refer to the same entity, such as `Colvin`, `S. C.`, and `Sidney Colvin`. See Gemini Notebook and The NLP Pipeline.
Epoch	One complete pass through the training set. Training usually takes many epochs. See Model Training.
F1 score	A single number combining precision and recall (their harmonic mean), from 0 to 1. It is low if either of them is low. See Evaluation Metrics.
Failure mode	A characteristic way in which a kind of system goes wrong. LLMs fail in patterns that follow from how they work, so you can learn to anticipate them. See Assessing Reliability and Failure Modes.
Fine-tuning	Continuing to train a pre-trained model on annotated examples from your own sources, so that it learns their particular names, spellings, and conventions. See Next Steps.
Generalization	Doing well on examples a model has never seen: the goal of training. See Model Training.
Generative AI	Systems that produce new content, such as text, images, audio, or code, instead of only classifying or predicting. It groups together systems with very different designs, from image generators to LLMs. See Basic Concepts.
Geocoding	Turning place names into coordinates, so they can be put on a map. Automatic geocoding needs checking: in the case study, a place where Stevenson lodged in San Francisco landed in Brooklyn. See Named Entities in Stevenson’s Letters.
GraphRAG	A variant of RAG that first uses an LLM to extract the entities and relationships in a collection into a network (a graph), and then uses the graph to answer questions about the collection as a whole. See Advantages of Programmatic Approaches.
Ground truth	The set of correct answers that a model’s output is compared with, usually made by people with the relevant expertise annotating a sample of documents. See Assessing Reliability.
Grounding	Anchoring a prompt in specific material: telling the model which documents to use, and asking it to show where each claim comes from. It improves transparency, not certainty. See Prompting Strategies.
Hallucination	The usual name for fluent, confident output that is false or invented, such as a biography of a person the model knows little about. See LLM Behaviour, and Failure Modes on the term itself.
Harmonic mean	A kind of average that is pulled down by the smaller of the values, so it penalizes a large gap between them. The F1 score is the harmonic mean of precision and recall. See Evaluation Metrics.
Inter-annotator agreement	How far two or more people annotating the same material agree. It tells you how reliable the ground truth itself is. See Ground-Truth Annotation.
Jupyter notebook	A document that mixes code, its output, and explanatory text. You run the code in small cells, one at a time, and see each result immediately below. See Setting Up Python.
Label	The correct answer attached to an example in supervised learning, such as the category of an image, or whether a name is a person, a place, or an organization. See Basic Concepts.
Language model	A model trained to predict the next token in a text. It generates text by repeating that prediction, one token at a time. See Language Models.
Large Language Model (LLM)	A deep learning model trained on very large collections of text, which generates text by predicting one token at a time. Chat assistants are built on LLMs. See Basic Concepts and Language Models.
Library	Code that someone else has written, which you install and import into your own, such as `pandas` for tables or `transformers` for models. See Setting Up Python.
Long Short-Term Memory (LSTM)	A kind of recurrent network with internal gates that decide what to keep and what to forget, so it can carry information further. LSTMs were the standard for translation in the mid-2010s. See Language Models.
Loss	A measure of how wrong a model’s prediction was. Training adjusts the weights, step by step, to reduce it. See Model Training.
Machine Learning (ML)	The part of AI in which systems learn patterns from examples, instead of following rules a programmer wrote. See Basic Concepts.
Model card	A page describing a published model: what it was trained on, which labels it predicts, how it was evaluated, and its known limitations. See The NLP Pipeline.
Named entity	A name in a text, of a type a model or an annotator is looking for. The case study uses three types: PER (people), LOC (places), and ORG (organizations). See Ground-Truth Annotation.
Named-Entity Recognition (NER)	The task of finding and classifying names in a text: people, places, organizations, and sometimes dates, works, or other categories. See The NLP Pipeline.
Natural Language Processing (NLP)	Computational methods for working with human language. See The NLP Pipeline.
Neural network	A mathematical structure, loosely inspired by the brain, made of simple connected units called neurons. The neurons are arranged in layers: data enters at the input layer, passes through one or more hidden layers, and emerges at the output layer. See Basic Concepts.
Open model	A model whose weights are published, so anyone can download it and run it on their own computer, where it never changes unless you change it. See Basic Concepts and Advantages of Programmatic Approaches.
Overfitting	When a model learns its training data too well, noise and quirks included, so it does well on that data and poorly on anything new. See also underfitting. See Model Training.
Pipeline	A fixed series of steps that turns raw material into results, such as loading, preprocessing, running a model, and analyzing its output. Each step can introduce errors, and errors compound. See The NLP Pipeline.
Pre-trained model	A model that someone else has already trained, and published for others to use. You download it and run it on your own text, without training anything. See The NLP Pipeline.
Precision	Of everything the model labelled as positive, the fraction that really was. Low precision means many false alarms. See Evaluation Metrics.
Preprocessing	Preparing text before a model reads it, such as converting ALL-CAPS words to title case. A mismatch between your text and a model’s training data can cause failures that look mysterious. See The NLP Pipeline.
Prompt	The text you give a model. The model’s actual input also includes a system prompt, documents, and the chat history. See Prompt Processing.
Prompt injection	Input crafted to make a model ignore or override its instructions. In indirect prompt injection, the text comes from somewhere else, such as a web page or a document the model is asked to read. See Hack Merlin.
Recall	Of everything that really was positive, the fraction the model found. Low recall means many misses. See Evaluation Metrics.
Recurrent Neural Network (RNN)	A neural network that reads text one token at a time, carrying a running summary forward. RNNs struggled to connect words that were far apart. See also LSTM and transformer. See Language Models.
Reinforcement Learning from Human Feedback (RLHF)	A form of reinforcement learning in which people compare a model’s outputs, and the model is tuned towards the answers they prefer. Much of the behaviour of chat assistants, including their refusals, comes from this stage. See Basic Concepts.
Reinforcement learning	Learning by trial and error: each output is scored, rewarded or penalized, and the model adjusts to earn more reward. See Basic Concepts.
Reproducibility	Getting the same result from the same input, with the same settings and the same versions, so that others (and you, later) can check and build on it. See Advantages of Programmatic Approaches.
Retrieval-Augmented Generation (RAG)	Searching your documents (or the web) for passages related to a question, and adding them to the model’s input. The model sees only the passages the search found, not the whole document. See Prompt Processing.
Sampling	Choosing the next token at random, weighted by the probabilities the model gives each one, instead of always taking the most likely. It’s why the same prompt can produce different answers. See Language Models.
Semi-supervised learning	Learning from a small amount of labelled data together with a large amount of unlabelled data. It is useful when labelling is slow or expensive. See Basic Concepts.
Sequence-to-sequence	A task that turns one sequence of tokens into another, such as translation. See Language Models.
Supervised learning	Learning from examples that come with the correct answers (labels), as if with an answer key. The NER models in Unit 3 were trained this way. See Basic Concepts.
Surface form	A name exactly as it’s written in a text, such as `Colvin`, `S. C.`, or `MY DEAR COLVIN`. NER models find surface forms, and entity resolution maps the different surface forms of one person or place to a single name. See Named Entities in Stevenson’s Letters.
Sycophancy	A model’s tendency to agree with you, flatter you, or give way when you push back. See Assessing Reliability and Failure Modes.
System prompt	Hidden instructions from the application, added to the model’s input, that say what role to play and how to behave. See Prompt Processing and Hack Merlin.
Temperature	A setting that controls how adventurous sampling is. At low temperature, the most likely tokens dominate, and at high temperature, less likely ones are chosen more often. See Language Models.
Test set	Data the model never sees during training or tuning, used once at the end to measure how well it performs on new examples. See Model Training.
Token	The unit a language model reads and writes: a whole word, part of a word, or punctuation. Rare names, ALL-CAPS text, and old spellings are often split into several tokens. See Language Models.
Tokenizer	The part of a language model’s system that splits text into tokens, from a fixed vocabulary. See Language Models.
Traceback	The error message Python prints when something goes wrong, listing the steps that led to the error. Read the last line first: it usually names the problem. See Coding with LLMs.
Training set	The part of the data a model learns from. See also development set and test set. See Model Training.
Transfer learning	Starting from a model that has already been trained on a large dataset, and training only part of it on your own examples. It’s why Teachable Machine needs only a few dozen examples. See Teachable Machine.
Transformer	The neural network design behind modern language models (introduced in 2017), which uses attention to process a whole passage at once. Not to be confused with `transformers`, the Python library used in Unit 3. See Language Models.
True and false positives and negatives	The four outcomes of a prediction. Positive and negative are what the model predicted (positive being the class you’re looking for), and true and false are whether it was right. A false positive is a false alarm, and a false negative is a miss. See Evaluation Metrics.
Tuning	Adjusting the settings of training, such as how large the model is or how long to train it, guided by results on the development set. See Model Training.
Underfitting	When a model is too simple to capture the real pattern, so it does poorly everywhere. See also overfitting. See Model Training.
Unsupervised learning	Learning from data without labels: the model looks for structure on its own, and groups similar items together. It’s up to you to work out what the groups mean. See Basic Concepts.
Vibe-coding	Running whatever code an LLM produces, and iterating until it seems to work, without understanding the code in between. See Coding with LLMs.
Virtual environment	An isolated set of Python libraries for one project, so that it doesn’t interfere with anything else on your computer. See Setting Up Python and Coding with LLMs.
Watermark	A hidden statistical pattern, added to AI-generated text by nudging the choice of tokens with a secret key. Usually only the provider can check for it, and heavy editing weakens it. See LLM Behaviour.
Weight	A number on each connection in a neural network, which determines how the value passing along it changes. The weights are what a network learns in training, and everything it knows is stored in them. See Basic Concepts.