Model words
158 terms. The words of language models, training and evaluation, in plain language.
A
- Active parameters
- The parameters a mixture-of-experts model actually uses for each token, usually a small share of its total.
- Agent
- A model that works towards a goal over many steps, choosing and using tools and reacting to the results.
- AI safety
- The work of preventing harm from AI systems, from everyday misuse to severe risks from very capable models.
- Alignment
- Making AI systems pursue the goals and values their designers and users intend, and avoid harmful behaviour.
- Attention
- The step in a transformer where each token weighs every other token and pulls in information from the ones that matter.
- Autoregressive generation
- Producing text one token at a time, feeding each new token back in as input for the next.
B
- Base model
- A model straight out of pre-training, which continues text but has not yet been tuned to follow instructions.
- Benchmark
- A fixed set of test questions with a scoring method, used to compare models on one kind of skill.
- Benchmark saturation
- When top models score so close to the maximum on a benchmark that it can no longer tell them apart.
- BERT
- Google's 2018 encoder model, which read text in both directions and set new records across language tasks.
- Byte-pair encoding
- A method for building a vocabulary by repeatedly merging the most frequent pair of symbols in a body of text.
C
- Calibration
- Whether a model's confidence matches how often it is right.
- Chain of thought
- Having a model write out intermediate reasoning steps before its answer, which improves accuracy on hard problems.
- Chatbot
- An application in which people talk with a model in turns, such as ChatGPT, Claude, Gemini or Meta AI.
- Chatbot Arena
- A public site where people compare two anonymous models' answers, used to rank chat models by preference.
- Chinchilla scaling
- DeepMind's 2022 finding that, for a fixed budget, models should be trained on about 20 tokens per parameter.
- Code-switching
- Mixing two or more languages in one conversation or sentence, as in Hinglish, a common challenge for models.
- Coding agent
- An agent that reads a codebase, edits files, runs commands and tests, and iterates until a change works.
- Computer use
- A model operating a computer as a person would, looking at the screen and moving the mouse and keyboard.
- Constitutional AI
- Anthropic's method of training a model to critique and improve its own answers against a written list of principles.
- Context window
- The maximum number of tokens a model can take in at once, counting the prompt, any documents and its own reply.
- Continuous batching
- Serving many users at once by adding and removing requests from a running batch at every generation step.
- Cross-entropy loss
- The standard training loss for language models: how surprised the model was by the token that actually came next.
D
- Data contamination
- When benchmark questions or answers end up in a model's training data, inflating its scores.
- Data filtering
- Selecting and cleaning training text, removing spam, toxic content, personal data and low-quality pages.
- Decoder
- A model, or part of one, that writes text one token at a time, each token seeing only those before it.
- Diffusion language model
- A model that writes by starting from noise or masked text and refining the whole passage over several steps.
- Direct preference optimisation
- A simpler alternative to RLHF that trains directly on preference pairs, with no separate reward model.
E
- ELMo
- A 2018 model from the Allen Institute for AI that produced contextual word embeddings from a two-way recurrent network.
- Embedding
- A list of numbers that stands for a token, word or passage, placed so that similar meanings sit close together.
- Embedding model
- A model that turns a passage of text into a single vector, used for search, clustering and retrieval.
- Emergent abilities
- Skills that seem to appear suddenly once a model passes a certain size, and the debate over whether they are real.
- Encoder
- A model, or part of one, that reads a whole text at once in both directions to build a representation of its meaning.
- Encoder-decoder
- A design with two stacks: one reads the input, the other writes the output while looking back at it.
- Evaluation
- Testing what a model can do and how it behaves, with benchmarks, human ratings and targeted test sets.
F
- Feed-forward network
- The part of each transformer layer that processes every token on its own, holding much of the model's stored knowledge.
- Few-shot prompting
- Putting a handful of worked examples in the prompt so the model copies the pattern.
- Fine-tuning
- Further training of an already trained model on a smaller, focused dataset to adapt it to a task or style.
- FlashAttention
- A way of computing exact attention that avoids slow trips to GPU memory, making long contexts faster and cheaper.
- Foundation model
- A large model trained on broad data that can be adapted to many tasks. A wider term than language model.
- Frontier model
- One of the most capable models at a given time, usually from a handful of well-funded labs.
- Function calling
- An API feature in which a model returns a structured request to run a named function with given arguments.
G
- Generative pre-trained transformer
- OpenAI's name for its series of decoder models, which has become a general label for models of this kind.
- GGUF
- A single-file format for quantised models, used by llama.cpp and other tools to run models on ordinary computers.
- GPQA
- A 2023 benchmark of very hard science questions written by PhD experts so that web searching does not help.
- Gradient descent
- The basic training rule: move every parameter a small step in the direction that lowers the loss.
- Graphics processing unit
- The chip that trains and runs language models, built for doing many simple calculations at once.
- Greedy decoding
- Always picking the single most likely next token. Fast and repeatable, but prone to dull or looping text.
- Grounding
- Tying a model's answer to specific sources, such as documents or search results, rather than its memory alone.
- Group relative policy optimisation
- A reinforcement learning method from DeepSeek that scores each answer against a group of answers to the same question.
- Grouped-query attention
- A form of attention in which several query heads share one set of keys and values, shrinking the KV cache.
- Guardrails
- Checks around a model that screen inputs and outputs for harmful, private or off-topic content.
H
- Hallucination
- When a model states something false or made up, such as a fake citation, with apparent confidence.
- Hidden state
- The vector a model holds for each token between layers, carrying everything it has worked out so far.
- Hugging Face Hub
- The largest public site for sharing model weights, datasets and demos, run by the company Hugging Face.
- HumanEval
- A 2021 coding benchmark of 164 Python problems, each checked by hidden unit tests.
- Hybrid model
- A model that mixes attention layers with state-space or recurrent layers to get the strengths of both.
I
- In-context learning
- A model picking up a new task from examples or instructions in the prompt, without any change to its weights.
- Indic language model
- A model built for the languages of India, such as Hindi, Bengali, Tamil, Telugu and Marathi.
- Inference
- Running a trained model to get answers, as opposed to training it. Most of the world's AI computing is now inference.
- Instruction tuning
- Fine-tuning a model on many tasks phrased as instructions, so it learns to do what it is asked.
- Interpretability
- Research into what happens inside a model: which parts represent which ideas, and how they produce an answer.
J
- Jailbreak
- A prompt crafted to get around a model's safety training and make it produce what it would normally refuse.
K
- Knowledge cutoff
- The date after which a model has seen no training data, so it knows nothing of later events unless told.
- Knowledge distillation
- Training a smaller student model to imitate a larger teacher model's outputs.
- KV cache
- Stored keys and values from earlier tokens, kept so the model does not recompute them for every new token.
L
- Language model
- Any system that assigns probabilities to sequences of words, and so can predict or generate text.
- Large language model
- A neural network trained on very large amounts of text to predict the next token, and so to read and write language.
- Latency
- How long a model takes to respond: the wait for the first token and the speed of the tokens after it.
- Layer
- One repeated block of a model. Text passes through dozens of layers, each refining the representation a little.
- Layer normalisation
- A step that rescales each token's vector to a steady size inside every layer, keeping training stable.
- Licence
- The legal terms under which a model's weights may be used, changed and shared.
- Logits
- The raw scores a model gives every token in its vocabulary before they are turned into probabilities.
- Long context
- The ability to read and use very long inputs, from hundreds of thousands to millions of tokens.
- Lost in the middle
- The tendency of models to use information at the start and end of a long input better than information in the middle.
- Low-rank adaptation
- A cheap fine-tuning method that freezes the model and trains small add-on matrices instead.
- Low-resource language
- A language with little digital text available for training, so models handle it less well.
M
- Masked language model
- A model trained to fill in hidden words in a sentence using the words on both sides, as BERT was.
- Mixed precision
- Training with lower-precision numbers for most calculations to save memory and time, while keeping key values precise.
- Mixture of experts
- A design in which each layer holds many expert sub-networks and a router sends each token to just a few of them.
- MMLU
- A widely used 2020 benchmark of about 16,000 multiple-choice questions across 57 subjects.
- Model API
- A web service that lets programs send prompts to a hosted model and get answers back, usually paid per token.
- Model card
- A short document published with a model describing what it is for, how it was trained and evaluated, and its limits.
- Model collapse
- The loss of quality and variety when models are trained, generation after generation, on text written by earlier models.
- Model Context Protocol
- An open standard from Anthropic for connecting models and agents to tools and data sources.
- Model specification
- A published document setting out how a lab intends its models to behave, and how to weigh conflicting goals.
- Multi-head attention
- Running several attention operations side by side, so different heads can track different kinds of relationship.
- Multi-head latent attention
- DeepSeek's attention design that compresses keys and values into a small shared vector to save memory.
- Multilingual model
- A model trained on many languages so it can understand and write in each, and often translate between them.
- Multimodal model
- A model that takes in, and sometimes produces, more than one kind of data: text, images, audio or video.
N
- Needle in a haystack test
- A long-context test that hides one fact in a long text and asks the model to find it.
- Neural network
- A computing system made of layers of simple units whose connection strengths are learned from data.
- Next-token prediction
- The training task of guessing the next token of a text from all the tokens before it.
O
- On-device AI
- Running a model on the user's own phone, laptop or car rather than in a data centre.
- Open source AI
- A model released with its weights, code and enough detail on its data that others can study and rebuild it.
- Open weights
- A model whose trained weights are published for anyone to download and run, whatever the licence terms.
P
- Parameter
- One adjustable number inside a model, learned in training. Model size is usually given as a parameter count.
- Perplexity
- A measure of how well a model predicts a text. Lower means the model was less surprised.
- Positional encoding
- Information added to token embeddings so the model knows the order of words, which attention alone ignores.
- Post-training
- Everything done to a model after pre-training: instruction tuning, preference tuning, safety training and reasoning training.
- Pre-training
- The first and largest stage of training, in which a model learns from trillions of tokens of general text.
- Process reward model
- A model that scores each step of a solution, not just the final answer, to guide reasoning.
- Prompt
- The text, and any images or files, given to a model to respond to.
- Prompt caching
- Saving the processed form of a long, repeated prompt so later requests that start the same way are cheaper and faster.
- Prompt engineering
- Writing and testing prompts to get better, more reliable answers from a model.
- Prompt injection
- Hiding instructions in content a model reads, such as a web page or email, so that it follows them instead of its user.
Q
- QLoRA
- LoRA fine-tuning on top of a 4-bit quantised model, which let large models be tuned on a single GPU.
- Quantisation
- Storing a model's weights with fewer bits, such as 8 or 4 instead of 16, to cut memory and speed up inference.
R
- Reasoning model
- A model trained to think at length, in a hidden or visible chain of thought, before giving its final answer.
- Recurrent neural network
- A network that reads a sequence one step at a time, carrying a memory forward. The main language approach before transformers.
- Red-teaming
- Deliberately attacking a model to find ways it can be made to misbehave, before others find them.
- Refusal
- When a model declines a request, ideally only for genuinely harmful ones.
- Reinforcement learning
- Learning by trial and reward rather than from fixed examples. Used to tune chat models and to train reasoning models.
- Reinforcement learning from human feedback
- Training a model with rewards learned from people's comparisons of its answers, the method that made ChatGPT helpful.
- Residual connection
- A shortcut that adds a layer's input to its output, so signals and gradients can pass through very deep networks.
- Retrieval-augmented generation
- Searching a collection of documents for relevant passages and giving them to the model along with the question.
- Reward hacking
- A model finding a way to score well on its training reward without doing what the reward was meant to encourage.
- Reward model
- A model trained to score answers the way human raters would, used to guide reinforcement learning.
- Rotary position embedding
- A way of encoding word order by rotating query and key vectors by an angle that depends on position.
S
- Sampling
- Choosing each next token from the model's probabilities, with some randomness, rather than always taking the top one.
- Scaling laws
- Regular patterns showing that loss falls predictably as models, data and compute grow.
- Self-attention
- Attention in which the tokens of one text attend to each other, rather than to a separate input.
- Self-consistency
- Asking a model the same question several times and taking the most common answer.
- SentencePiece
- A tokenizer library from Google that treats text as a raw stream, spaces included, so it works for languages written without spaces.
- Sliding-window attention
- Attention limited to a fixed number of recent tokens, so cost stops growing with the full length of the text.
- Small language model
- A compact model, usually under about 10 billion parameters, that can run on a laptop, a phone or modest servers.
- Softmax
- A function that turns a list of raw scores into probabilities that are all positive and add up to one.
- Sovereign AI
- A country building its own models and computing capacity, so it does not depend entirely on foreign providers.
- Special token
- A reserved token that marks structure rather than text, such as the end of a document or the start of a user's turn.
- Speculative decoding
- Speeding up generation by letting a small fast model draft several tokens and the big model check them in one pass.
- Speech model
- A model that turns speech into text, text into speech, or holds a spoken conversation directly.
- State-space model
- A family of sequence models that carry a fixed-size running state instead of attending to every past token.
- Structured output
- Making a model reply in an exact machine-readable format, such as JSON matching a given schema.
- Supervised fine-tuning
- Fine-tuning on example prompts paired with good answers, usually the first step of turning a base model into an assistant.
- SWE-bench
- A benchmark of real bug reports from open-source Python projects, where a model must produce a fix that passes the tests.
- Sycophancy
- A model's tendency to tell users what they want to hear, agreeing with them even when they are wrong.
- Synthetic data
- Training examples written by other models rather than collected from people, now a large share of post-training data.
- System card
- A detailed report a lab publishes with a major model, describing its safety testing, risks and safeguards.
- System prompt
- Hidden instructions placed before the conversation that set a model's role, rules and tone.
T
- T5
- Google's 2019 encoder-decoder model that framed every language task as turning one text into another.
- Temperature
- A setting that controls how adventurous sampling is. Low temperature gives safe, predictable text; high gives varied text.
- Test-time compute
- Spending more computation while answering, for example by thinking longer, to get better results.
- Token
- The unit of text a model reads and writes: a whole word, a piece of a word, a punctuation mark or a byte.
- Tokenizer
- The program that splits text into tokens and turns them into numbers before a model sees them, and back again afterwards.
- Tokenizer fertility
- How many tokens a tokenizer needs, on average, for one word of a given language. Lower is cheaper and faster.
- Tool use
- A model asking an outside program, such as a calculator, search engine or database, to do part of a task.
- Top-p sampling
- Sampling only from the smallest group of tokens whose probabilities add up to a set share, such as 90 per cent.
- Training compute
- The total amount of computation used to train a model, usually counted in floating-point operations.
- Training data
- The text, code and other material a model learns from, now measured in trillions of tokens.
- Transformer
- The neural network design, introduced in 2017, that almost every modern language model is built on.
V
- Vector database
- A database built to store embeddings and quickly find those nearest to a query vector.
- Vision encoder
- The part of a multimodal model that turns an image into vectors, usually by cutting it into small patches.
- Vision-language model
- A model that reads images together with text, answering questions about photos, documents and charts.
- Vocabulary
- The fixed list of tokens a model knows, each with its own number and its own embedding.
W
- Watermarking
- Hiding a statistical signal in generated text so that it can later be identified as machine-written.
- Weights
- The full set of learned parameters of a model, saved as files. Having the weights means being able to run the model.
- Word2vec
- A 2013 Google method for learning word embeddings from which much of modern language modelling grew.
Z
- Zero-shot
- Asking a model to do a task with no examples, just an instruction.