RESOURCE GLOSSARY

AI Training Data Glossary

Plain-language definitions of the terms used to source, evaluate, license and prepare AI datasets. Browse by topic to clarify requirements and compare supplier specifications.

Browse Glossary by Topic

How to Use This Glossary

Each entry explains a term, not a guarantee about a particular dataset. Confirm deliverables and rights in the project specification and agreement.

For purchasing steps, see the Dataset Buyer's Guide and its AI Dataset Quality Checklist.

Core AI Training Data Terms

AI training data: Data used to fit or fine-tune a model. Validation and test data serve separate development and assessment roles.
Dataset: An organized collection of records or files, with supporting information, used for training, evaluation or analysis.
Generative AI training data: Data used to train or fine-tune models that generate or transform text, images, video, audio or designs. Separate evaluation data measures their outputs. See Generative AI Training Data.
Training set: The portion of data used to fit model parameters and learn patterns or relationships.
Validation set: Data used during development to select settings or model versions. Keep it separate from training data and the final test set.
Test set: Data reserved for assessing a model after training and tuning. Its separation from development data must be established, not assumed from its name.
Fine-tuning data: Task- or domain-specific data used to adapt an already trained model to new requirements.
Evaluation dataset: Data used to measure model performance on defined tasks, conditions or cases. The evaluation protocol determines reference data and separation requirements. See Model Evaluation Datasets.
Benchmark dataset: A fixed or versioned test collection used with a defined task, scoring protocol and references where needed to compare model performance.

Dataset Type Terms

Image dataset: A collection of still images used for tasks such as classification, object recognition, visual search, computer vision, multimodal AI, or model evaluation. See Image Datasets for AI Training.
Video dataset: A collection of video files, clips, or extracted frames used for video understanding, motion analysis, activity recognition, generative video, or temporal AI tasks. See Video Datasets for AI Training.
Audio dataset: A collection of speech, environmental sounds, object sounds, music, or other audio recordings prepared for audio classification, generation, transcription, source separation, or model evaluation.
Audio-visual dataset: A multimodal dataset that connects video with synchronized audio using shared files, timestamps, event boundaries, or alignment metadata.
Document dataset: A collection of PDFs, scanned documents, forms, reports, or other document files used for OCR, extraction, classification, retrieval, layout understanding, or Document AI. See Document Datasets for AI Training.
Template dataset: A collection of editable design layouts, presentation templates, or structured creative assets used for design automation, layout understanding, component prediction, or generative design workflows. See Template Datasets.
Text dataset: Language records for NLP, extraction, retrieval, summarization or instruction tuning. The focus is text and its labels, rather than document page structure. See Text Datasets for AI Training.
Multimodal dataset: Data linking two or more modalities, such as image-text or video-audio. IDs, timestamps or alignment fields record how the inputs relate. See Multimodal AI Datasets.
Image-text dataset: A dataset that connects images with captions, descriptions, labels, questions, answers, or other text fields for vision-language training, retrieval, captioning, or evaluation.
Video-text dataset: A dataset that connects video clips or frames with captions, descriptions, questions, answers, or metadata for video understanding, retrieval, captioning, or vision-language tasks.
Human-authored data: Content written by people, such as documents, captions or answers. Record any machine assistance separately; human review of generated text is a different production method.
Synthetic data: Data generated artificially to represent required patterns, events, conditions, or edge cases rather than captured directly from the target real-world event. Its generation method and intended use should be documented.
Custom dataset: Data produced or curated to a buyer's task, coverage, format, annotation, quality and rights requirements. See Custom Dataset Creation.

Computer Vision and Visual AI Terms

Computer vision: AI methods for interpreting images and video, including objects, scenes, people, actions and their visual relationships. See Computer Vision Datasets.
Object recognition: Identifying an object category or specific instance in visual data, such as a tool type or product variant. See Object Recognition Datasets.
Object detection: Identifying and locating objects in an image or video frame, typically using class labels and bounding boxes.
Image classification: The task of assigning one or more predefined category labels to an image.
Image segmentation: The task of dividing an image or frame into labeled regions or object instances, commonly using pixel-level or polygon-based masks.
Scene understanding: Interpreting a visual scene, including its environment, objects, people, activities and relationships.
Human activity recognition: The task of identifying actions, movement, gestures, routines, or behavior categories in images, videos, or frame sequences. See Human Activity Recognition Datasets.
Object tracking: The task of maintaining an object's identity and location across consecutive video frames, commonly using frame-level boxes, masks, points, or trajectories.
Facial expression analysis: Analysis of visible facial movements or expression categories. These cues do not establish a person's internal emotional state. See Facial Expression Analysis Datasets.
Avatar and presenter AI workflows: Tasks involving talking heads or virtual presenters, including speech-related facial motion, gestures, posture, gaze and framing. See Avatar & Presenter AI Datasets.

Language, Multimodal, and VLM Terms

Vision-language model (VLM): A model that processes visual information and language together for tasks such as captioning, visual questions or retrieval. See Vision-Language Model Datasets.
VLM dataset: Paired visual and language data prepared for VLM training or evaluation, with task-specific captions, questions, instructions or grounding references.
Captioned dataset: A dataset of images, videos, audio, documents, or other media paired with descriptive text, usually linked through stable identifiers or file-level records.
Visual question answering: A task where a model receives visual content and a related question, then produces or selects an answer based on the visible information.
Cross-modal retrieval: The task of finding relevant content across different modalities, such as retrieving images with text queries or matching video clips to descriptions.
Natural language processing (NLP): Computational methods for processing human language, including classification, extraction, generation and summarization. Document AI can combine NLP with OCR and layout analysis.
Large language model dataset: Data prepared for LLM pretraining, fine-tuning or evaluation. A retrieval corpus may support an LLM application without being used to train the model.
Instruction-tuning data: Instructions or prompts paired with target responses to train a model to follow directions. Records may also contain context, constraints or quality labels.
Retrieval corpus for RAG: Searchable source material retrieved to inform a generated response. Using a corpus for retrieval does not by itself update model parameters.

Document AI and OCR Terms

Optical character recognition (OCR): The process of detecting and extracting text from document images, scanned pages, photographs, screenshots, or other visual files.
Document AI: AI workflows that read, classify, extract, retrieve, or understand information from documents. See Document Datasets for AI Training.
Document classification: The task of assigning a document to one or more predefined classes based on its text, layout, visual appearance, metadata, or combined features.
Form extraction: The task of identifying and extracting structured fields, values, selections, or relationships from forms.
Table extraction: The task of detecting and extracting rows, columns, cells, headers, values, and relationships from tables in documents.
Layout understanding: The task of interpreting document structure, including text blocks, tables, headers, footers, forms, sections, fields, and visual hierarchy.
Reading order: The intended sequence in which text blocks, fields, tables, captions, and other document regions should be read or processed.
Extracted text: Text obtained from a document, image, or video frame through OCR, manual transcription, document parsing, or another extraction workflow.

Annotation and Metadata Terms

Data annotation: Task-specific reference information added to data, such as labels, boxes, masks, keypoints, captions, time boundaries or extracted fields.
Metadata: Information describing files, their technical properties, origin, rights or versions. Metadata supports management and selection but is not necessarily a training label.
Dataset label: A task-defined category, state or target assigned to a record, object, frame, action, document or field.
Caption: A text description connected to an image, video, frame, audio file, document, or other dataset item.
Keyword tag: A descriptive search or organization term. A tag is not automatically a controlled class label or verified annotation.
Taxonomy: A structured category system used to organize classes, labels, tags, or metadata fields, often through parent-child relationships.
Ontology: A formal representation of entities, properties, relationships, and rules used to describe how concepts in a dataset relate to one another.
Ground truth: An accepted reference label, answer or measurement used for training or evaluation. Its reliability depends on how it was collected and checked.
Bounding box: A rectangular annotation marking the location and approximate extent of an object or region within an image or video frame.
Segmentation mask: A pixel-level or region-based annotation identifying the area occupied by a class, object instance, foreground subject, or other target region.
Keypoint annotation: A set of labeled points marking defined locations such as body joints, facial landmarks, object parts, or other reference positions.
Temporal annotation: An annotation that identifies when an event, action, state, sound, shot, or other segment begins and ends within time-based data.
Quality assurance (QA): Processes for checking files, annotations, metadata and documentation against agreed quality and acceptance criteria.

Licensing, Rights, and Trust Terms

Rights-cleared dataset: A dataset for which the provider has documented the rights and permissions supporting the specified licensed use. Rights-cleared does not mean unrestricted use; permitted AI activities, releases, limitations, and license terms must still be reviewed.
Dataset licensing: The contractual framework defining how a dataset may be used, including permitted AI activities, commercial scope, restrictions, documentation, and delivery terms. See Dataset Licensing & Compliance.
Commercial AI use: Using data in workflows connected to paid products, enterprise systems, commercial model development, customer-facing services, or other business applications, subject to the agreed license.
Model release: Documentation recording a person's permission for specified uses of their identifiable likeness. Whether a release covers a particular AI training or commercial use depends on its language and the accompanying dataset license. See Model & Property Releases.
Property release: Documentation granting specified permissions relating to identifiable private property, interiors, locations, artwork, or other protected assets. Its applicability to a particular AI use must be reviewed with the dataset license.
Likeness rights: Rights and legal interests relating to the use of a person's identifiable appearance, image, voice, or representation. Applicable requirements vary by jurisdiction and intended use.
Usage scope: The activities a buyer is permitted to perform with a dataset under an agreed license, such as training, fine-tuning, evaluation, benchmarking, internal research, or commercial product development.
Dataset provenance: Records of data origin, creation, acquisition and transformation that connect delivered assets to their sources. See Dataset Provenance & Source Traceability.
Chain of rights: The documented sequence showing how relevant rights or licensing authority pass from creators or rights holders through the provider to the permitted buyer use.
Data exclusivity: A contractual restriction on licensing the same data to other parties. Its scope may be limited by dataset, use case, customer, industry, geography, model type, or time period.

Dataset Delivery and Operations Terms

Dataset delivery: Packaging and transferring dataset files and agreed documentation to the buyer. See Dataset Delivery & Secure Transfer.
Secure dataset delivery: An agreed workflow for transferring and handling dataset files using appropriate access controls, transfer methods, integrity checks, and buyer-specific security requirements.
Dataset package: The delivered files and agreed supporting materials, such as annotations, metadata, manifests, provenance records and usage terms.
Dataset manifest: An inventory of delivered files, commonly listing IDs, paths, formats, sizes, versions, checksums and links to metadata or split assignments.
Checksum: A value calculated from file contents and compared with a reference to detect changes. It does not itself encrypt files, control access or establish source authenticity.
Dataset versioning: Identifying dataset releases and recording changes to files, annotations, metadata, splits and documentation.
Sample dataset: A subset supplied for review under agreed sample terms. Confirm how it was selected and how well it represents the full collection.

Dataset Quality Terms

Dataset quality: How well data meets the intended task across relevance, coverage, integrity and annotation quality. Rights and delivery suitability need separate checks. See AI Dataset Quality Checklist.
Dataset coverage: The range and quantity of required classes, scenarios, subjects, conditions and edge cases present in a dataset.
Dataset diversity and variation: Variation across relevant dimensions such as subjects, environments, layouts, viewpoints, lighting, formats or languages.
Dataset representativeness: How closely the composition and distribution of data reflect the intended deployment or evaluation population and conditions.
Class balance: The distribution of examples across labels or categories. Appropriate balance depends on the model task and target environment and does not always require equal class sizes.
Dataset deduplication: Identifying and managing exact or near-duplicate records within or across datasets. Related derivatives may need grouping rather than deletion.
Data leakage and train-test contamination: Use of information during training or model selection that should be unavailable under the evaluation protocol. Train-test contamination includes overlapping test examples or close duplicates; shared subjects or sources matter when the protocol requires separation.
Annotation accuracy: How closely annotations match the accepted references and labeling rules for the task.
Hard negative: A non-target example that closely resembles the target class or commonly causes model errors, included to improve discrimination or test model robustness.
Dataset noise: Irrelevant, corrupted, ambiguous, low-quality, or incorrectly labeled data that reduces usefulness for the intended model task.
Edge case: An unusual, difficult, or less common example that remains relevant to the intended model behavior or deployment environment.

Buyer and Procurement Terms

Dataset brief: A written summary of the buyer's task, required content, volume, formats, labels, quality, rights and delivery needs. See Custom Dataset Creation.
Dataset requirements: The content, technical, annotation, quality, rights, security and delivery conditions the data must meet.
Enterprise AI dataset: Data packaged with the technical, commercial and supporting documentation needed for an organization's review and use. See Enterprise AI Datasets.
Dataset vendor: A supplier that licenses, curates, produces, annotates or delivers data. Confirm which roles it performs directly.
Procurement review: The buyer's assessment of price, supplier capability, rights, evidence, security, delivery and contract terms before purchase.

Use These Terms to Evaluate or Request Data

Apply these definitions to a buying decision, collection review or dataset brief.

Evaluate a Dataset

Use the purchasing guide to review specifications, sample evidence, licensing and quality criteria before approving a dataset.

Dataset Buyer's Guide

Browse Available Datasets

Explore licensed image, video, audio, motion and multimodal collections, then check their scope against your model requirements.

Browse Dataset Library

Request a Custom Dataset

Send your modality, content coverage, volume, annotation needs, rights requirements and delivery constraints for a scoped request.

Request Dataset