Appendix A is a glossary organizing common terms related to AI, machine learning, and LLMs, grouped for easy lookup. Find a term by its heading, and follow the arrow references to check related concepts. For this book's own methods of confirmation and its authority boundaries, see Appendix B.
AI and Machine Learning Basics
AI — Artificial Intelligence
AI is a broad field/concept covering technologies and mechanisms used to tackle intellectual tasks with computers and the like. Prediction, classification, and generation are examples of this, and machine learning is one representative method related to AI. AI as a whole is not the same thing as machine learning, generative AI, or LLM.
ML — Machine Learning
Machine learning is a method that uses data to adjust a model's parameters, producing a model useful for making predictions or generating output from input. It is used as part of AI — not all of AI is machine learning.
Deep Learning
Deep learning is a branch of machine learning that uses neural networks with multiple hidden layers. It is not the same thing as machine learning or AI as a whole, and not every neural network qualifies as deep learning. → Neural Network
Neural Network
A neural network is a model, or its structure, made up of interconnected computational units arranged in layers, which uses adjustable values such as weights to produce output from input. It is widely used in machine learning, but it is not literally a reproduction of a biological brain. Deep learning is a method that uses neural networks with multiple hidden layers. → Parameter, Deep Learning
Dataset
A dataset is a collection of data or examples sharing a common format or grouping. In machine learning, it may be used for training, validation, testing, or evaluating a model. Not every dataset has labels attached, and dataset, training data, test set, and labeled data are not the same thing. → Label / Labeled Data
Parameter
A parameter is a value internal to a model that affects how it computes its output. The weights of a neural network are an example, adjusted from data during training. This is distinguished from a hyperparameter, which is set before or between training runs. → Neural Network
Algorithm
An algorithm is a defined computational procedure or set of rules for carrying out a given task. In machine learning, the procedure for determining a model's parameters from data is called a machine learning algorithm. An algorithm is a procedure — it is not the resulting model itself. → Machine Learning, Parameter
Feature
A feature is input-side information, property, or representation used by a model or process in machine learning. It's not limited to using the original data as-is — it can include representations obtained through transformation. A feature is neither an internal model parameter nor the label being predicted. → Parameter, Label / Labeled Data
Label / Labeled Data
A label is the desired answer or output information paired with an example in supervised learning. Labeled data refers to data that combines such labels with examples. A label can be not just a classification category but also a numerical value predicted in regression, and not every dataset includes one. A feature is input-side information, distinguished from a label. → Feature, Classification, Regression
Classification
Classification is a machine learning task that predicts or assigns a class or category for a given input. The prediction target is discrete classes or categories, not a continuous value. This includes binary classification, multi-class classification, multi-label classification, and more — classification is not a term for one specific algorithm. → Regression
Regression
Regression is a machine learning task that predicts or estimates a continuous numerical value from input. It's distinguished from classification, which deals with classes or categories, and like classification, it refers to a kind of task, not one specific algorithm. → Classification
Models and Architecture
Foundation Model
A foundation model is a base model trained on broad data that can be adapted and reused for multiple downstream tasks or uses. It may be trained at large scale, but it is not a concept defined by scale alone. It's distinguished from a model built for one specific task, and it is not synonymous with LLMs or with large-scale models in general.
Layer
A layer is a single stage inside a neural network or model that bundles together a computational unit or transformation. Stacking layers transforms the representation of information step by step, from input toward output. A layer is not a Neuron itself, and its composition and number differ by model.
Neuron
A neuron is a computational unit (node) used in a neural network. It combines input values, using things like weights, to produce a transformed output passed on to the next stage of processing. It is not a literal reproduction of a biological neuron.
Model
In machine learning, a Model is a mathematical/computational construct, obtained through learning, used to produce predictions, output, or representations from input. It works based on a state or design adjusted through learning. A model is distinguished from an Algorithm (which is a procedure), from weights alone, and from a user-facing AI service itself. → Algorithm, Model Weights
Embedding
An Embedding represents a token, item, or object as a numerical vector that a model can work with. This representation is used in computation and comparison, and the vector space may reflect relationships useful to the model. It is not the original string or object itself, nor its meaning as such. → Feature
Vector Database
→ Embedding, Semantic Search, Retrieval
A Vector Database is a database that stores and indexes Vectors, such as Embeddings, and can be used for similarity-based search, among other things. A Vector Database is not RAG itself — it is one component that can be used in RAG or Semantic Search.
Vector
→ Embedding
A Vector is an ordered sequence of numbers. In machine learning, a Vector can represent various kinds of information, such as features, probabilities, or Embeddings. A Vector itself and an Embedding are not the same thing.
Transformer
A Transformer is a neural network architecture built primarily around Attention. It processes relationships between representations, and comes in variants such as Encoder-based, Decoder-based, and combined forms. It's used in many modern language models, but it is not the same thing as Attention itself, LLM itself, or every generative AI. → Attention, Encoder, Decoder
Attention
Attention is a mechanism that handles relationships between different positions or elements of an input or representation, changing the weight or contribution of each piece of information when producing an output or new representation. It's one important mechanism making up a Transformer, but it is not the entire Transformer itself. → Transformer
Model Weights
Model weights are the numerical values learned inside a model that affect how input is transformed and output is computed. In a neural network, they are an example of parameters adjusted during training. Weights alone do not necessarily constitute a complete model artifact, and they don't necessarily include the architecture, tokenizer, or execution environment either. → Parameter, Model, Checkpoint
MoE — Mixture of Experts
MoE (Mixture of Experts) is a model design pattern with multiple sub-networks or components acting as "experts." Depending on the input, a routing or gating mechanism changes which experts are used, or how much each contributes — so the model's components are used conditionally. It is not always the same thing as a conventional ensemble.
Checkpoint
A checkpoint is a saved state of a model or its training at a given point in time. It often includes the model's weights, and depending on the framework or workflow, may also include training-related state such as an optimizer's. A saved state can be used to resume, evaluate, or distribute — it's not necessarily the final published model, and it doesn't necessarily mean weights alone. → Model Weights
Activation Function
An Activation Function is a function that transforms the value computed by a unit or layer within a neural network's computation. Since the transformed value is passed on to the next stage of processing, it plays a role in introducing nonlinear transformation or behavior. The function itself is not a learned model parameter. → Neuron, Parameter
Encoder
An Encoder is a component that processes input and converts it into a representation used internally by the model — such as a representation that incorporates context. In a Transformer, the Encoder's layers process the input representation. Encoder is not the same thing as Decoder or the Transformer as a whole, and it is used in other machine learning configurations too. → Transformer, Decoder
Decoder
A Decoder is a component that uses a model's internal representation to produce an output, or the next representation of output. In a Transformer, the Decoder's layers process representations and are involved in generating the output sequence. Decoder is not the same thing as Encoder or the Transformer as a whole, and not every Decoder operates autoregressively or is an LLM. → Transformer, Encoder
Encoder-Decoder
→ Encoder, Decoder, Transformer
Encoder-Decoder is a configuration combining an Encoder, which converts input into an internal representation, and a Decoder, which generates/reconstructs output from that internal representation. Transformers come not only in Encoder-Decoder form, but also in Encoder-only and Decoder-only forms.
Language Models and Context
NLP — Natural Language Processing
Natural Language Processing (NLP) is a field concerned with processing and using human language on a computer. It covers tasks related to analyzing, processing, and generating language, but it is not a term that refers only to LLMs or generative AI. → Language Model, LLM
Language Model
A Language Model is a model that handles patterns or probabilities appearing in language or token sequences, predicting or estimating sequences or outputs related to language. Not every language model is large-scale, Transformer-based, or conversational. → LLM, Token, Next-token prediction
LLM — Large Language Model
A Large Language Model is a language model treated as large-scale in current technical usage, used for a broad range of language-related tasks. There's no fixed numerical threshold — by parameter count or data volume — that universally defines "large" here. LLM is not the same thing as AI as a whole, a Chat Model, a Foundation Model, or Transformer itself. → Language Model
Token
A Token is the unit a model uses when handling input such as text. Depending on the tokenizer and model, a token may correspond to a whole word, part of a word, a character, or a byte — so a token and a word are not necessarily the same. → Tokenization
Tokenization
Tokenization is the process of converting input such as text into tokens or token identifiers a model can work with. How the split happens, and at what stage, differs by tokenizer and model; tokenization is not the same thing as embedding or understanding meaning itself. → Token, Embedding
Vocabulary
→ Token, Tokenization
Vocabulary is the set of tokens or symbols handled by a language model or tokenizer, among other things. The units contained in a vocabulary don't necessarily match natural-language "words," and vocabulary is not the same thing as a token itself or the process of tokenization.
Context
Context is the information available to a model or system for the current processing or inference. It may include things like the current input, prompt, or conversation history, but what's included differs by system. Context is a separate concept from the context window (the upper limit on how much can be processed), memory (retained for later reuse), and training data itself. → Context Window, Prompt, Memory
Context Window
A context window is the range/capacity of context a model or system can handle in a single pass. This is sometimes expressed in token count, but the specific limit and behavior differ by model and system. It does not mean long-term memory or training data, and it does not mean that a larger window always produces better results. → Context, Memory, Token
Prompt
A prompt is input, an instruction, or other content given to guide a model's processing or output. It can include a question, an instruction, examples, or data, but a prompt is not necessarily the entirety of the available context. → Input, Context, Output
Input
Input is data or information given to a model or system for processing. Its form and range differ by model and task, and it doesn't refer only to text a user directly types. Input is not always the same thing as a prompt, which gives an instruction or content. → Prompt, Output
Output
Output is data, information, or a result a model or system returns as the result of processing its input. Its form differs by task and model — it can be text, a classification result, a numerical value, and more. Output does not always mean the final answer displayed to a user. → Input, Model, Classification, Regression
Next-token prediction
Next-token prediction is the process/objective of predicting the token that follows, based on preceding tokens and the available context. It's used by autoregressive language models and is relevant to many modern LLMs, but it does not describe the structure of every AI or every language model, and by itself it does not show human-like understanding. → Token, Language Model, LLM
Chat Model
A Chat Model is a model or system adapted/configured to conduct a conversational exchange. It may handle a dialogue involving multiple rounds of messages, but the specific message format and mechanism differ by model and system. Even if it's based on a pretrained language model, not every Chat Model is an LLM or shares a single architecture. → Language Model, LLM, Context
Memory
Memory refers to information or state retained/stored so it can be reused across later stages or interactions. The same information may be described as Context, from the perspective of what's usable in the current process, or as Memory, from the perspective of retaining and re-obtaining it for later use — the functional viewpoint differs. The mechanism differs by implementation, so Memory is not always the same thing as any one of Context, the Context Window, Model Weights, a Checkpoint, or Training Data. → Context, Context Window, Model Weights, Checkpoint
Learning and Adaptation
Training
Training is the process of adjusting a model's parameters or learned state, based on data and a learning objective. Training is distinguished from Inference, which uses a trained model to obtain predictions or output. Training as a whole is not the same thing as just Fine-tuning. → Parameter, Inference, Fine-tuning
Inference
Inference is using a trained model, by giving it input, to obtain predictions or output. Its role differs from Training, which adjusts a model's parameters, and it isn't a term that refers only to text generation. → Training, Input, Output
Pretraining
Pretraining is learning done prior to later adaptation to a task, domain, or instruction. It refers to a stage that precedes later Fine-tuning or Post-training, and it does not always take place through unsupervised learning, at web scale, or for the same objective. → Training, Fine-tuning, Post-training
Fine-tuning
Fine-tuning is performing additional learning or adaptation on a model that has already been trained, often a pretrained one. It may use data or an objective suited to a particular task, domain, or instruction, and the range of parameters updated differs by method. Fine-tuning is not always the same thing as Pretraining, all of Post-training, or Instruction Tuning. → Pretraining, Instruction Tuning, Post-training, Parameter
Supervised Learning
Supervised Learning is a method that learns from examples that pair input with a target/label or desired output to be predicted. That target information guides the model's learning. It's not limited to classification — it's also used for tasks like regression, which predicts numerical values. → Label / Labeled Data, Unsupervised Learning, Reinforcement Learning / RL
Reinforcement Learning / RL
Reinforcement Learning (RL) is a learning method in which an agent or policy interacts with an environment and improves its behavior while receiving reward or return signals. The reward is not necessarily given directly by a human, and RL is not the same thing as RLHF or supervised learning as a whole. → Reward Function, RLHF, Objective Function
RLHF — Reinforcement Learning from Human Feedback
RLHF is a method that uses human feedback — such as evaluations or preferences — in a reinforcement-learning-based adaptation of a model. Human feedback may be used as a reward signal or related learning information, but the specific stages and procedures differ by system. It is not the same thing as RL as a whole, all of alignment, all of Post-training, or Instruction Tuning. → Reinforcement Learning / RL, Reward Function, Instruction Tuning, Post-training
Reward Model / Preference Learning
→ Reward Function, RLHF, Alignment
A Reward Model is a model trained, using comparison/evaluation data obtained from humans and others, to predict how much a given output is preferred. Preference Learning refers to the broader idea of learning from this kind of preference information. A Reward Model is not always the same thing as a Reward Function itself, and Preference Learning does not mean only RLHF.
Loss Function
A Loss Function is, in learning or optimization, a function that computes a quantity representing a learning-related discrepancy, such as prediction error. That quantity is normally treated as something to be minimized during training. A Loss Function is not always the same as a Reward Function or an evaluation metric, and its terminology can overlap with an Objective Function depending on context. → Objective Function, Reward Function, Training
Reward Function
A Reward Function is, in reinforcement learning, a function or rule that determines what reward is assigned to actions, states, or outcomes. The reward value or signal that results from, or is used through, this mechanism is distinguished from the function itself. The reward is information that guides the improvement of a learned policy or behavior, and it is not always the same as human feedback, a Loss Function, or an Objective Function. → Reinforcement Learning / RL, RLHF, Loss Function
Objective Function
An Objective Function is a quantity or function that optimization or learning tries to minimize or maximize. Depending on context, it may include multiple elements, such as loss, reward, or regularization. It is not always the same as a Loss Function or a Reward Function, and technical documents sometimes use these terms with overlapping meaning. → Loss Function, Reward Function, Training
Hyperparameter
A Hyperparameter is a setting that shapes a model's structure or the course of learning/optimization, separate from a Parameter, which is adjusted directly through training. It's often decided before a training run, though in some cases it may be tuned automatically. It's distinguished from a Parameter, which is obtained through training. → Parameter, Training
Unsupervised Learning
Unsupervised Learning is a learning method that finds patterns or structure in data lacking the explicit target labels used in supervised learning. This does not mean there is no goal or criterion at all, and it is not always the same thing as Self-supervised Learning or Pretraining. → Supervised Learning, Self-supervised Learning, Pretraining
Instruction Tuning
Instruction Tuning is a method that adapts a model using data such as instructions and their corresponding responses, aiming to improve its ability to follow instructions. It's sometimes treated as a form/use of Fine-tuning, but it does not mean every kind of Fine-tuning, Post-training, or RLHF, and it does not necessarily require reinforcement learning. → Fine-tuning, RLHF, Post-training
SFT / Supervised Fine-Tuning
→ Fine-tuning, Supervised Learning, Instruction Tuning
SFT (Supervised Fine-Tuning) is a method of Fine-tuning a pretrained model using supervised data, such as input paired with a desired output. It's sometimes used as part of Instruction Tuning, but it is not the same thing as every kind of Fine-tuning or RLHF.
Self-supervised Learning
Self-supervised Learning is a method that creates its own learning target or signal from the data itself, rather than from labels separately prepared by a person. Even when no external target label is given, it uses a target constructed from the data, so it doesn't mean "there is no label or target at all." It is not always the same as Unsupervised Learning or all of Pretraining. → Unsupervised Learning, Supervised Learning, Pretraining
Post-training
Post-training is a stage/grouping term for the additional learning or adaptation done after Pretraining. It can include things such as Fine-tuning, Instruction Tuning, preference-based adaptation, and RLHF, but which methods are included differs by system and context. Post-training is not one single technique, RLHF, or every kind of Fine-tuning. → Pretraining, Fine-tuning, Instruction Tuning, RLHF
Distillation
Distillation is a method that trains/adapts a student model using information — such as output, probabilities, internal representations, or behavior — shown by a teacher model. It may aim at knowledge transfer between models or greater efficiency, but it doesn't mean the student fully reproduces the teacher's information, or that it's always done with the same method or goal. It is not always the same method as Fine-tuning. → Model, Model Weights, Fine-tuning
Generative AI and Reliability
Generative AI
Generative AI is a field/category of AI systems or models that generate content — text, images, audio, video, code, and more — based on input or context. The range of content or modality it handles differs by context, and it does not mean all of AI or LLM. → AI, LLM, Generation
Generation
Generation is the process by which a model produces content or output, based on input, context, the model's state, and more. It is not the field/category of Generative AI itself, and not every Inference operation is called Generation. → Generative AI, Input, Output, Inference
Hallucination
Hallucination is a term for a case where a model's generated output, though it may look plausible, is incorrect, unsupported, or inconsistent with the relevant evidence or context. The exact boundary — how much weight is given to error, lack of support, or claims about reality — varies slightly by source and material, and it is not the same thing as every model error or intentional deception. → Factuality, Grounding, Output
Multimodal
Multimodal describes a model or system that handles two or more modalities (kinds of information handled) — such as text, images, audio, or video — as input or output. The combination of input and output, and which modalities are supported, differs by system, and being multimodal does not imply generation capability. → Input, Output, Generative AI
VLM / Vision-Language Model
→ Multimodal, Model
A VLM (Vision-Language Model) is a model that handles visual information, such as images, together with language information. Uses include answering questions about an image or describing an image, but it is not always the same thing as multimodal AI as a whole.
Bias
Bias is a term for a systematic tendency or skew appearing in data, predictions, output, or evaluation. It can refer to a statistical/technical deviation in one context, and to a social stereotype or fairness issue in another — the meaning differs by context. So Bias does not always refer only to unfairness, or only to training data. → Dataset, Evaluation, Factuality
Nondeterminism / Stochasticity
Nondeterminism / Stochasticity is a term related to the property that the same input is not guaranteed to produce the same output, and to the probabilistic mechanisms behind that. Stochasticity may particularly emphasize probabilistic choice or processing, while Nondeterminism may emphasize that reproducing a result is not guaranteed. It expresses the possibility that output may vary — it does not mean the output is always different, or that it cannot be trusted. → Sampling / Decoding, Output
Alignment
Alignment is an effort or property aimed at making a model's or system's behavior better match intended goals, instructions, preferences, constraints, and values. Adaptation using human feedback is one method, but Alignment itself is not always the same thing as RLHF, safety, correctness, or mere obedience. → RLHF, Instruction Tuning, Model
Grounding
Grounding is connecting a generated output or response to, and basing it on, given, retrieved, or otherwise available information/evidence/context. This can help suppress ungrounded generation, but it does not guarantee factual correctness, and it does not mean only RAG or displaying citations. → Context, Factuality, Hallucination
Confidence / Confidence Score
Confidence / Confidence Score is a value or score representing an estimate of certainty or likelihood in a given task or model/system. Its meaning depends on what was measured and how, and it does not always represent a calibrated probability across every system, nor does it guarantee correctness. It's not always the same as a natural-language self-report of "I'm confident," either. → Factuality, Evaluation, Output
LLM-as-a-judge
LLM-as-a-judge is a method that uses an LLM to evaluate, compare, score, or classify the output of a model or the result of a task. Its judgment may approximate agreement with human preference, but it is not an objective ground truth or a human evaluation itself, and it can be affected by biases related to prompting, presentation order, explanation length, or self-preference — so it is not reliable by definition. → LLM, Evaluation, Bias
Contamination / Benchmark Contamination
Contamination / Benchmark Contamination is a state in which benchmark or test material, or information closely related to it, ends up included in a model's Training or adaptation before evaluation. This can compromise the independence of the evaluation and make measured performance look higher than it actually is, but it is not the same thing as every kind of overfitting or data leakage, and it isn't necessarily intentional. → Training, Evaluation
Factuality
Factuality is, in the context of evaluation, the property/quality perspective of how well the claims contained in an output correspond to confirmable facts or reality. The criteria differ by evaluation method and task; it does not represent the fluency of the output or an "intent to convey the truth," and it does not simply mean that Grounding was performed. → Hallucination, Grounding, Confidence / Confidence Score
Sampling / Decoding
Sampling / Decoding is the procedure that selects an output token or sequence, during generation, from the model's scores or probability distribution. Depending on the source, terminology can overlap — for example, treating Decoding as the broader name for selection strategies and Sampling as one probabilistic form of selection — but Sampling does not cover every form of Decoding, and neither is Training itself. → Generation, Token, Inference
Temperature
→ Sampling / Decoding, Nondeterminism / Stochasticity
Temperature is a setting value that adjusts how the probability distribution is treated during generation, affecting the variability of the output. Generally, increasing it raises diversity, and decreasing it makes selection more concentrated, but Temperature alone does not determine the content or reproducibility of the output.
Search, Retrieval, and RAG
Search
Search is a process/feature that takes a query against a collection of information or documents and finds information or results relevant to the goal. Web search is one example of Search, but Search does not always target only the web, and it is not the same thing as Generation or Retrieval. → Retrieval, Generation
Retrieval
Retrieval is finding and taking out documents or information relevant to an information need, from a collection or index. It is closely related to search, but it is not Generation itself, and Retrieval alone does not mean RAG as a whole. → Search, RAG, Generation
RAG (Retrieval-Augmented Generation)
RAG is a method/configuration that obtains relevant information from an external source and uses that information to perform generation. It's an approach that combines Retrieval and Generation, but using RAG does not guarantee the factuality of the output, and it is not the same thing as retraining the model. → Retrieval, Generation, Grounding, Factuality
Knowledge Base
→ Retrieval, RAG
A Knowledge Base is an expression for a collection of information organized so it can be used for search or reference. In generative AI, it may be used as an information source for RAG or Retrieval, but a Knowledge Base doesn't necessarily consist of a Vector Database, nor does it mean the model's internal knowledge itself.
Training Data
Training Data is the data used to Train a model or machine learning system. It's a separate concept from the trained model itself and from Current Context, and is also distinguished from test data, which is kept separate for evaluation. → Training, Dataset, Parameter, Context
Synthetic Data
Synthetic Data is data artificially generated by simulation, statistical methods, generative models, and the like — not data directly collected or observed from the real world. Synthetic Data is not necessarily generated by AI, and it's not necessarily used as Training Data either; being synthetic alone does not guarantee quality, anonymity, or freedom from bias. → Dataset, Training Data, Generative AI
Semantic Search
Semantic Search is search that finds relevant results by considering the meaning/context of the query and the target information, not just matching strings or keywords. Some implementations use Embeddings or vector search, but Semantic Search does not always mean the same thing as vector search, and it is not RAG itself. → Search, Embedding, RAG
Chunk / Chunking
A Chunk is a portion/unit obtained by splitting a document or data into fixed ranges. The process of splitting it this way is called Chunking. It may be used in processes such as search, Retrieval, Embedding, and RAG, but a Chunk is a different unit from a Token, and Chunking is not Tokenization or RAG itself. The size and method of splitting also differ by use case. → Token, Tokenization, Retrieval, RAG
Tool Use and Agents
Tool Use
Tool Use is the mechanism/capability by which an AI model or AI system uses external tools — search, APIs, functions, code execution, and more — to obtain information, process it, or perform operations. A model generating a request to use a tool and an external tool actually being executed are not the same thing, and Tool Use does not mean only Function Calling. → Function Calling, AI Agent
AI Agent
An AI Agent is an AI system that, for a given goal or task, judges the situation using a model and the like, and — as needed — combines multiple steps and Tool Use to carry out processing. Not every LLM application or chatbot is an AI Agent, and being an AI Agent does not mean full autonomy. → Tool Use, LLM, Agentic AI
System Prompt
A System Prompt is an instruction given to a model — kept distinct from ordinary user input — that specifies the model's behavior, role, constraints, response policy, and the like. Its name, precedence, and handling may differ by system or API, and a System Prompt does not mean the entire policy or security boundary of the system itself. → Prompt, Context
Function Calling
Function Calling is a mechanism in which a model outputs, in structured form, a function name and arguments in order to use a predefined function or tool. That output is a request or instruction to execute the function — it does not by itself mean the function was actually executed. → Tool Use, Output
Agentic AI
Agentic AI is an expression for an AI system, or a property of an AI, that combines situational planning, judgment, Tool Use, and operations to a certain degree of autonomy in pursuit of a goal. This term is used with some range in meaning — it is not always exactly the same as AI Agent, and it does not necessarily mean full autonomy or the use of multiple agents. → AI Agent, Tool Use
Inference Server
An Inference Server is a server or piece of software that accepts Inference requests to a model, executes the target model, and returns results. An Inference Server is neither the model itself nor the process of Inference itself, and it does not always mean the entire AI service either. → Model, Inference
Evaluation and Performance
Evaluation
Evaluation is examining a model's or AI system's performance or behavior using data, tasks, criteria, metrics, human judgment, and the like, suited to the purpose. There are various methods of Evaluation — it does not mean only using a Benchmark, or a single metric. → Benchmark, Test Set
Benchmark
A Benchmark is an evaluation framework — a combination of shared tasks, data, evaluation methods, and metrics — used to evaluate and compare the performance of models or systems. A Benchmark's result shows performance under specific evaluation conditions; by itself it does not represent every real-world capability or quality. → Evaluation, Test Set
Robustness
Robustness is a property indicating how well a model or system can maintain its intended performance or behavior even when there is variation in input, environment, or conditions. What Robustness is being evaluated against differs by context, and Robustness does not mean correctness, safety, and generalization all at once. → Evaluation, Generalization
Ground Truth
Ground Truth is the actual observed result, correct label, reference value, or the like, treated as the comparison standard when evaluating a model's predictions or output. Information treated as Ground Truth can also include measurement error, recording mistakes, or differences in human judgment — it does not mean a complete, absolute truth in every situation. → Evaluation, Test Set
Generalization
Generalization is a model's property of working appropriately, for its intended purpose, on new data it did not directly see during training — not merely fitting the individual examples of the Training Data. Generalization is not the memorization of Training Data, and it does not mean that performance is guaranteed on every unknown situation or differently distributed data. → Training Data, Overfitting, Test Set
Overfitting
Overfitting is a state in which a model fits Training Data excessively and cannot generalize sufficiently to new data. It can show up as high performance on Training Data but degraded performance on unseen data, but not every case of poor performance is called Overfitting, and the cause is not limited to just one thing. → Training Data, Generalization, Test Set
Reproducibility
Reproducibility is a term for the property that, when an experiment or evaluation is repeated under stated conditions and methods, sufficiently consistent results are obtained. Different fields distinguish terms like Reproducibility and Replicability differently, and it does not mean that output will match character-for-character under the same conditions, or that AI's output is always deterministic. → Evaluation, Nondeterminism / Stochasticity
Test Set
A Test Set is a collection of data set aside — not used in a model's Training — to evaluate the trained model. Its role differs from Training Data, and a Test Set itself is not always the same thing as a Benchmark or Ground Truth. How data is split and the evaluation procedure differ by use case. → Training Data, Evaluation, Benchmark
Underfitting
Underfitting is a state in which a model fails to sufficiently capture important patterns or structure in the data it's learning from, and cannot achieve sufficient performance even on the Training Data. There are various possible causes — a model's representational capacity, its features, the learning method, and the amount of learning, among others — it does not simply mean short training time or a small model. → Training, Training Data, Generalization, Overfitting