Skip to content

Concept Glossary: T

This page lists English glossary entries for this letter. Entry bodies are generated from per-term source files with includes.

target

  • Meaning: A target is the answer column or goal value that a supervised learning model is meant to predict. It marks which part of a table is the answer for the current learning task.
  • Why it matters: A table does not automatically define the learning problem. Readers need to separate input features from the target so the model does not learn from the answer column or from information too close to it.
  • Related concepts: feature, dataset, supervised learning label, task definition, target candidate
  • Core Section: P1-8.1
  • Appears in: P2-12.3, P3-1.1, P3-1.2, P3-2.2, P3-3.2, P3-9.1, P3-9.2, P3-9.3, P3-9.9, P3-9.11, P3-9.12, P4-2.1, P4-10.1

true objective

  • Meaning: The true objective is the real goal that a reward, metric, or proxy is trying to stand in for. In reinforcement-learning applications, it may mean long-term satisfaction, safety, or stability behind easier numbers such as clicks or speed.
  • Why it matters: If the reward number and the true objective differ, a system can improve what is easy to measure while harming the actual goal. This concept anchors reward-design and reward-hacking checks.
  • Related concepts: reward design, proxy target, reward hacking, objective function
  • Core Section: P4-19.3
  • Appears in:

task definition

  • Meaning: Task definition turns a real-world goal into the input, output, data, and evaluation criteria that a model can actually work with.
  • Why it matters: The same broad goal can become classification, retrieval, summarization, recommendation, or another task depending on the output and evaluation boundary. Clear task definition keeps model choice, dataset design, and success criteria from drifting.
  • Related concepts: model, model input, model output, dataset, model evaluation design
  • Core Section: P1-4.4
  • Appears in: P3-1.1, P3-2.2, P7-index, P7-1.1, P7-summary

tensor

  • Meaning: A tensor is a general name for a multidimensional numeric array used to hold inputs, intermediate representations, and outputs in deep learning. It includes scalars, vectors, and matrices as simpler cases.
  • Why it matters: Images, sentences, audio, and tabular batches look different on the surface, but learning frameworks usually compute with them as tensors. The concept lets readers interpret model inputs, batches, hidden representations, and outputs by shape and axis.
  • Related concepts: scalar, vector, matrix
  • Core Section: P5-9.2
  • Appears in: P2-1.1, P2-1.2, P5-9.1, P5-10.1, P5-summary

test data

  • Meaning: Test data is data held out from model training and used to check how the trained model behaves on data treated as unseen. It is an evaluation sample, not proof of all real-world performance.
  • Why it matters: Test data is a basic guard against judging a model only by data it already used for learning or tuning. It helps readers treat reported scores as estimates that depend on the test set design.
  • Related concepts: training data, sample, generalization, model validation
  • Core Section: P2-5.3
  • Appears in: P2-6.2, P2-15.2, P4-4.1, P4-4.2, P4-5.1, P4-5.2, P4-7.3

text and data mining in the training-data context

  • Meaning: Text and data mining in the training-data context means analyzing large amounts of text or data to find patterns, statistics, and relationships. In debates about AI training data, it is connected to the question of when reading and analyzing material is allowed.
  • Why it matters: Some regions discuss text and data mining through research or data-analysis exceptions, but that does not mean every use of every material is automatically allowed. This concept helps separate model training, data analysis, and redistribution of original material.
  • Related concepts: training data, copyright, quotation
  • Core Section: P1-15.2
  • Appears in:

threshold

  • Meaning: A threshold is a cutoff value used to turn an output number into an operational action such as automatic handling, human review, or an alert.
  • Why it matters: The model score or comparison value is not the same as the rule that turns it into action. Moving the threshold can change false alarms, missed cases, automation scope, and human review load without retraining the model.
  • Related concepts: probability estimate, probability calibration, classification
  • Core Section: P1-7.3
  • Appears in: P1-7.2, P3-2.2, P3-5.7, P3-6.2, P3-9.8, P3-9.12, P4-6.1, P4-6.2, P4-6.4, P4-11.1, P4-11.2, P4-14.1, P4-15.3, P7-7.4

time split

  • Meaning: A time split is a data split that keeps earlier observations on the training side and later observations on the validation or test side when time order matters. It avoids randomly mixing records that would not have been available together at prediction time.
  • Why it matters: Randomly splitting time-ordered data can make performance look better by letting future patterns influence training. A time split asks whether only information available before the later case was used, which helps expose leakage and evaluation-design mistakes.
  • Related concepts: data leakage, evaluation design, prediction contract
  • Core Section: P3-9.13
  • Appears in: P3-9.13, P4-4.1

token

  • Meaning: A token is the basic computational unit into which a model splits text for processing. It is not always the same as a word: one word can split into several tokens, and several short expressions can sometimes be handled as one piece. A sentence that looks continuous to a person becomes a sequence of tokens inside the model.
  • Why it matters: Readers need this distinction because the units people read and the units a model computes over can differ. Cost, context-window limits, and token coverage all begin from this computational unit. Understanding tokens also explains why prompt design and context limits must be judged by token count rather than by character count alone.
  • Related concepts: next-token prediction, language modeling, embedding, tokenization, context window
  • Core Section: P6-2.1
  • Appears in: P1-10.2, P5-13.1, P5-13.2, P6-2.1, P6-2.2, P6-2.3, P6-2.4, P6-2.5, P7-4.1, P7-4.2

tokenization

  • Meaning: Tokenization turns raw text into a sequence of tokens a model can process. A sentence that looks whole to a person becomes an ordered series of token pieces for the model.
  • Why it matters: Tokenization affects token count, cost, context use, chunk boundaries, and how inputs are interpreted. In P3-6.2 the same idea is used by analogy: segment tokens turn a raw curve into a shorter sequence so its order and direction can be compared.
  • Related concepts: token, context window, Byte Pair Encoding
  • Core Section: P6-2.2
  • Appears in: P3-6.2, P6-2.2, P6-2.3, P6-2.4, P6-2.5, P7-4.1, P7-4.2, P7-5.1

tool use

  • Meaning: Tool use is the structure where a model or AI agent calls an external capability for lookup, calculation, storage, modification, or transmission.
  • Why it matters: Reading information and causing an external system to act carry different responsibilities and risks. Tool use makes permission, approval, failure recovery, and human oversight part of the AI-system design.
  • Related concepts: AI agent, permission, least privilege, prompt injection
  • Core Section: P6-14.1
  • Appears in: P1-14.2, P1-14.3, P1-14.4, P1-14.5, P1-14.6, P6-10.3, P6-14.2, P6-15.1, P7-6.1, P7-6.2

training data

  • Meaning: Training data is the data a model directly uses to learn patterns or rules. In supervised learning, it may contain inputs and answers; in generative AI, it may refer to large-scale source material used during training.
  • Why it matters: What a model learned from affects performance, bias, generalization, copyright questions, and privacy questions. The term helps readers separate data used for learning from data held out for evaluation.
  • Related concepts: dataset, test data, generalization, bias
  • Core Section: P1-15.2
  • Appears in: P2-5.3, P2-6.2, P2-15.2, P4-1.2, P4-4.1, P4-4.2, P4-5.1, P4-5.2, P4-7.3, P4-12.1

Transformer

  • Meaning: A Transformer is a family of neural network architectures that uses attention so multiple positions in a sequence can refer to one another without recurrence. Instead of passing information only step by step from front to back, it compares relationships among positions in parallel and builds contextual representations. Because token meaning alone does not encode order, Transformers also supply position information separately.
  • Why it matters: Transformers are a core structural basis of modern LLMs, but they need to be read together with pretraining, scaling, and deployment choices. This concept helps explain that an LLM is not merely a large model; it is the result of combining attention-based architecture with large-scale learning strategies. It also clarifies why language modeling shifted away from RNN-style sequential processing toward architectures that can compare long-range context more directly.
  • Related concepts: Attention, language modeling, direct lineage
  • Core Section: P1-11.3
  • Appears in: P1-9.3, P5-13.2, P5-14.1, P5-14.2, P5-15.1, P6-4.1, P6-4.2, P6-4.3, P7-5.1

transparency

  • Meaning: Transparency means making visible where AI is involved, what purpose it serves, and what limits apply. It does not require exposing every internal calculation, but it does require showing the boundaries needed for interpretation and challenge.
  • Why it matters: Users need to know whether a result involved AI, what evidence or limits apply, and where review is possible. Transparency supports accountability and oversight by making automated and human roles easier to distinguish.
  • Related concepts: accountability, human oversight, supporting evidence, provenance
  • Core Section: P1-15.1

transformative use

  • Meaning: A use that adds new purpose, meaning, or character rather than merely substituting for the original.
  • Why it matters: It is important in copyright discussion, but must be judged in context rather than assumed automatically.
  • Related concepts: Read together with the surrounding glossary entries and the linked Sections for this term.
  • Core Section: P1-15.2

trajectory

  • Meaning: An ordered path of states, actions, observations, or positions over time.
  • Why it matters: It is useful for reinforcement learning, robotics, and sequence analysis because the order of events changes the interpretation.
  • Related concepts: Read together with the surrounding glossary entries and the linked Sections for this term.
  • Core Section: P1-7.4

trust boundary

  • Meaning: A boundary between parts of a system that should not automatically trust each other.
  • Why it matters: It is crucial when prompts, retrieved text, tools, credentials, and user input meet inside one AI workflow.
  • Related concepts: Read together with the surrounding glossary entries and the linked Sections for this term.
  • Core Section: P1-14.4