Glossary

AI training data glossary, in plain English

This AI training data glossary collects the 89 terms that come up on real projects, from bounding boxes to word error rate. Each one gets a sentence or two, and a link to the page where it matters.

Browse the terms Read the guides

Labeling and quality
Speech, audio and conversation
Text and languages
Images and video
Testing a model

speech

Word error rate (WER) The share of words a speech model gets wrong against a checked transcript.

term 1

images

Bounding box A rectangle drawn around an object to show where it is.

term 2

labeling

Inter-annotator agreement How often two people label the same item the same way.

A B C I S W

term 3

85 89 terms each in a sentence or two

✓ Linked to the source see each term in context

Labeling and quality

The words that come up when people agree on what a label means, and check that everyone applied it the same way.

Annotation guideline: The written rules annotators follow: the label set, examples for each label, and decisions on the awkward cases. Most label quality problems start here.; More on Data annotation
Entity tagging: Marking spans of text as people, places, products, dates or other types. You'll also see it called named entity recognition labeling, or NER.; More on Data annotation
Inter-annotator agreement: How often two annotators give the same item the same label. Low agreement usually points to an unclear guideline, not careless people.; More on Data annotation
Label set: Every class an item can get, including what counts as "other". Some teams call it the taxonomy or the ontology.; More on Data annotation, Text data
Segmentation: Marking the exact outline or pixels of an object or region, instead of a rectangle around it. It takes longer and gives a model more detail to learn from.; More on Data annotation
Sentiment labeling: Marking whether a piece of text reads as positive, negative or neutral. Some projects add finer grades, or a separate label for sarcasm.; More on Data annotation

Speech, audio and conversation

Recording, transcribing and labeling what people say, from a read prompt to a messy two-sided call.

Barge-in: When a caller starts talking before the system has finished speaking. Good voice bots have to handle it.; More on Conversational AI data
Clean transcription: The same speech with the fillers and false starts taken out, so it reads like writing. Also called non-verbatim.; More on Transcription
Intent: What the speaker wants done, like checking a balance or canceling an order. A bot's first job is to recognize it.; More on Conversational AI data, Text data
IVR: Interactive voice response: the phone menu that answers a call and routes it, by keypad or by speech.; More on Conversational AI data
Read speech: Speech recorded from prompts, so the words are known in advance and every one you need gets covered.; More on Speech data
Scripted dialogue: Conversation where speakers follow prompts, so the intents and wording you need are covered.; More on Conversational AI data
Signal-to-noise ratio: How loud the speech is compared with the background. Noisy audio has a low ratio.; More on Speech data
Slot: A detail the bot needs to finish the task, like a date, an amount or an account type. Also called an entity.; More on Conversational AI data
Sound tags: Marks for sounds that aren't speech, such as [noise], [laughter] or [silence], written the same way every time.; More on Transcription
Speaker diarization: Working out who spoke when, and marking each turn with a speaker label.; More on Transcription
Speaker metadata: Details recorded with each clip, like language, accent, age group, place and device.; More on Speech data
Speech recognition (ASR): Software that turns spoken words into text. It learns from audio paired with checked transcripts.; More on Speech data
Spontaneous dialogue: Unscripted conversation about a task, with the hesitation, repetition and corrections of real speech.; More on Conversational AI data
Spontaneous speech: Natural, unscripted talk, with the pauses, fillers and corrections of real conversation.; More on Speech data
Timestamp: A time code that ties a stretch of text to its place in the audio, per line, per segment or per word.; More on Transcription
Turn: One speaker's go in a conversation, from when they start talking to when the other side replies.; More on Conversational AI data
Utterance: One stretch of speech from one speaker, usually a sentence or a short request.; More on Conversational AI data, Speech data
Verbatim transcription: Every word as it was said, with fillers, repeats, false starts and stutters kept in.; More on Transcription
Word error rate (WER): The share of words a speech model gets wrong against a reference transcript, counting words swapped, dropped and added. A reference with its own mistakes makes the number meaningless.; More on Transcription, Speech data

Text and languages

Written data, translation, and the things that make Indian languages their own problem.

Code-mixing: Using two languages in one message, like Hindi and English in the same sentence.; More on Text data
Code-switching: Moving between languages inside one conversation or sentence, like Hindi and English in one line.; More on Indian language data, Transcription, Speech data
Corpus: A collection of text gathered for a purpose, like training or testing a language model.; More on Text data
Dialect: A regional form of a language, with its own sound and some of its own words.; More on Indian language data
Localization: Adapting text for a market beyond the words: dates, units, names, formality and tone.; More on Translation and localization
Low-resource language: A language with little public data available for training AI models.; More on Indian language data, Translation and localization
Machine translation post-editing: A person correcting a machine's draft until it's right. Also called MTPE.; More on Translation and localization
Named entity: A specific thing mentioned in the text, like a person, product, place, date or amount.; More on Text data
Native speaker review: A check by someone who grew up speaking the language, who catches what a learner misses.; More on Indian language data
Natural language processing (NLP): The field of AI that reads, understands or writes human language.; More on Text data
Placeholder: A variable like {name} or {date} in a string, filled in by software later. It has to survive translation untouched.; More on Translation and localization
Register: How formal the language is. Many languages mark it in grammar, like the formal and informal forms of "you".; More on Translation and localization
Romanized text: An Indian language typed in Latin letters, common on phone keyboards, with spelling that varies by person.; More on Indian language data
Scheduled languages: The 22 languages listed in the Eighth Schedule of India's constitution.; More on Indian language data
Script: The writing system a language uses, like Devanagari for Hindi or Tamil script for Tamil.; More on Indian language data
Sentiment: Whether a piece of text reads as positive, negative or neutral.; More on Text data
Term base: The agreed list of key terms and their translations, so every translator uses the same words.; More on Translation and localization
Text expansion: Translated text often runs longer than the source, so interface text has to be checked for fit.; More on Translation and localization
Transliteration: Writing a word in another script while keeping its sound, like Hindi written in Latin letters.; More on Indian language data, Transcription, Translation and localization, Text data

Images and video

Boxes, outlines, frames and tracks: how a picture or a clip gets turned into something a model can learn from.

Action recognition: A model telling what's happening in a clip, like walking, waving or crossing.; More on Video data
Bounding box: A rectangle drawn around an object in a photo, with the object's class attached.; More on Image data, Computer vision, Data annotation
Capture brief: The written plan for new photos: subjects, places, light, angles and devices.; More on Image data
Class: One of the named categories an object can belong to, like car, person or traffic sign.; More on Computer vision
Class balance: How evenly the classes are represented. A class with few photos is one a model learns badly.; More on Image data
Edge case: A rare or awkward photo, like odd light or a strange angle, that a model still has to handle.; More on Image data, Data annotation, Synthetic data
Egocentric video: Video filmed from the point of view of the person doing the activity, by a camera they wear on the head, chest or glasses.; More on Egocentric video data
Exocentric video: Video filmed by a separate camera watching the person, also called third-person video.; More on Egocentric video data
Frame: One still image in a video. A clip is many frames shown one after another.; More on Video data
Frame rate: How many frames a video has per second. It changes how finely movement is captured.; More on Video data
Hand-object interaction: The moments a hand reaches for, holds, uses or lets go of an object, labeled so a model learns how things get handled.; More on Egocentric video data
Image classification: Giving the whole photo one label from a set, like plant, fruit or bottle.; More on Image data
Instance segmentation: A separate outline for each object, so two cars get two masks.; More on Image data, Computer vision
Key frame: A frame chosen for labeling when not every frame needs labels.; More on Video data
Keystep: One step of a longer task, like crack the egg or tighten the bolt, labeled with where it starts and ends.; More on Egocentric video data
Object tracking: Following the same object across video frames and keeping one id for it the whole time.; More on Computer vision, Video data
Occlusion: When part of an object is hidden behind something else. Labeling rules say how to handle it.; More on Image data, Computer vision, Video data
Polygon annotation: An outline made of many points that follows an object's shape more closely than a box.; More on Computer vision
Semantic segmentation: Every pixel in the photo assigned to a class, so all the road is road.; More on Image data, Computer vision
Temporal segment: A stretch of video with a start and an end, labeled with what happens in it.; More on Video data
Track id: The number that stays with one object for the whole clip, like car 03.; More on Video data
Truncation: When an object is cut off by the edge of the image, so only part of it is in the frame.; More on Computer vision

Testing a model

How you find out whether a model actually works, and for whom.

Benchmark: A fixed test the model runs against, with agreed correct answers and target scores, so results can be compared over time.; More on Data and model evaluation
Data leakage: When test items also appear in the training data. The model has seen the answers, and the scores look better than they are.; More on Data and model evaluation
Ground truth: The answers checked by people that model outputs get compared with. If the ground truth is wrong, so is every score.; More on Data and model evaluation
Human evaluation: People rating model outputs against agreed rules, used where there's no single correct answer to compare with.; More on Data and model evaluation
Model drift: A slow drop in quality after launch, because the language, products or users have changed since the model was trained.; More on Data and model evaluation
Precision and recall: Precision is how many of the model's answers were right. Recall is how many of the right answers it found.; More on Data and model evaluation
Subgroup analysis: Splitting results by group, such as age, accent or language, to find where a model underperforms.; More on Data and model evaluation
Test set: Data kept aside to measure the model, never used to train it. It should look like the people who'll use the model.; More on Data and model evaluation

LLM data

Prompts, answers and human judgments: the data that shapes how a large language model behaves.

Instruction tuning: Training a model on prompts paired with good answers, so it learns to follow instructions. Also called supervised fine-tuning.; More on LLM training data
Large language model (LLM): A model trained on large amounts of text to read and write language, like the ones behind chat assistants.; More on LLM training data
Multi-turn conversation: A dialogue with several back-and-forth turns, used to teach a model to keep track of a thread.; More on LLM training data
Preference data: Records of which answer people preferred when shown two or more answers to the same prompt.; More on LLM training data
Red-teaming: Writing difficult or adversarial prompts on purpose, to find where a model gives unsafe or wrong answers.; More on LLM training data
Reward model: A model trained on preference data to score answers the way people would, used during RLHF.; More on LLM training data
RLHF: Reinforcement learning from human feedback: tuning a model with human judgments of which answers are better.; More on LLM training data
Rubric: The written criteria raters use to score an answer, like accuracy, tone and whether it followed the prompt.; More on LLM training data

Synthetic data

Generated data, and the checks that keep it honest.

Data augmentation: Making new training examples by changing existing ones, like flipping an image or adding noise to audio.; More on Synthetic data
Domain gap: The difference between generated data and the real world. A model trained only on synthetic data can stumble on it.; More on Synthetic data
Mixed training set: Real and synthetic data used together, usually with real data kept back for testing.; More on Synthetic data
Privacy leakage: When generated data reveals details from the real records it was based on.; More on Synthetic data
Realism review: People checking whether generated data looks, reads or sounds like the real thing.; More on Synthetic data
Seed data: A small set of real examples that generated data is modeled on. How it's used affects privacy.; More on Synthetic data
Synthetic data: Data generated to look and behave like real data, instead of being collected from real people or events.; More on Synthetic data

Missing a term?

Tell us which one and we'll add it. Krishsinh reads every message and replies within one business day.

Talk to our team