NVIDIA NCA-GENM key terms

A neural network architecture that processes sequences using stacked self-attention and feed-forward layers, enabling parallel training and strong modeling of long-range dependencies in text, audio, and images.

An attention mechanism where each token computes relevance scores against every other token in a sequence, producing context-aware representations; it is the core operation inside transformer blocks.

The process of splitting raw text into smaller units called tokens, such as words, subwords, or characters, before mapping them to integer IDs for a language model.

A dense numeric vector that represents a token, sentence, image, or other item in a continuous space where similar meanings are close together and usable by machine learning models.

A transformer-based neural network trained on massive text corpora to predict and generate sequences, enabling tasks such as summarization, translation, question answering, and code generation.

Further training a pretrained model on a smaller, task-specific dataset so its weights adapt to a particular domain or objective, usually improving accuracy over prompting alone.

A parameter-efficient fine-tuning method that freezes the base model and trains small low-rank matrices inserted into selected layers, drastically reducing trainable parameters and memory needs.

A training pipeline that uses human preference comparisons to train a reward model, then optimizes a language model with reinforcement learning to produce outputs people prefer.

A model design where a gating network routes each input token to a small subset of specialized sub-networks, increasing total capacity while keeping per-token compute relatively low.

The practice of designing instructions, context, examples, and output constraints in a model input to elicit accurate, relevant, and safe responses without changing model weights.

A pattern that retrieves relevant documents from an external knowledge store and inserts them into the prompt so a language model can answer with fresher, grounded context.

Reducing the numerical precision of model weights or activations, such as from FP16 to INT8 or FP8, to lower memory use and speed inference, sometimes with small accuracy loss.

Stored key and value tensors from previous attention steps that let an autoregressive model avoid recomputing past tokens, trading memory for faster sequential text generation.

A distributed strategy that splits individual model layers and their matrix operations across multiple GPUs, enabling models too large for one device to run with coordinated computation.

Processing multiple inputs together in one model invocation to improve hardware utilization and throughput, though batch size affects latency and memory consumption.

NVIDIA's parallel computing platform and programming model that lets developers run general-purpose workloads on GPUs, forming the foundation for many deep learning frameworks.

An NVIDIA SDK for optimizing and deploying trained neural networks, performing layer fusion, precision calibration, and kernel selection to reduce inference latency on NVIDIA GPUs.

An end-to-end framework for building, training, and customizing generative AI models such as LLMs, speech, and multimodal systems, with tools for data curation and distributed training.

A set of containerized microservices that expose optimized pretrained and custom models through standard APIs, simplifying secure deployment of generative AI applications.

A change in input data distribution over time compared with training data, which can degrade model performance and trigger retraining or monitoring alerts.

A resampling technique that splits data into multiple folds, training on some folds and validating on others, to estimate model performance and reduce evaluation variance.

A controlled experiment comparing two variants, such as model versions or prompts, by randomly assigning users or samples and measuring a predefined outcome metric.

A result is statistically significant when the observed effect is unlikely to arise by chance under a null hypothesis, typically judged with a p-value threshold and confidence intervals.

Creating modified copies of training examples through transformations such as cropping, rotation, paraphrasing, or noise injection, to improve robustness and reduce overfitting without collecting new data.

A measure of similarity between two vectors computed as the cosine of their angle, ranging from -1 to 1 and commonly used to compare embeddings in retrieval systems.

A storage system built to index and query high-dimensional embeddings, supporting fast nearest-neighbor search for semantic retrieval, recommendation, and retrieval-augmented generation workloads.

Precision measures how many retrieved or predicted items are relevant, while recall measures how many relevant items were found; both are useful for evaluating classification and retrieval.

Machine learning that combines and relates multiple data modalities, such as text, images, audio, and video, to improve understanding or generation beyond what a single modality provides.

The process of combining features or outputs from different modalities into a shared representation or decision, using strategies such as early fusion, late fusion, or cross-attention.

A transformer architecture that splits an image into patches, embeds them as tokens, and applies self-attention to classify or analyze visual content, often pretrained on large image datasets.

A self-supervised training approach that pulls representations of related pairs, such as an image and its caption, closer together while pushing unrelated pairs apart.

The automated conversion of text in scanned documents or images into machine-readable characters, often used as a preprocessing step in document understanding pipelines.

Technology that transcribes spoken audio into text by modeling acoustic features and language patterns, enabling voice interfaces, captions, and multimodal applications.

A model output that sounds fluent and confident but contains fabricated or unsupported information, often caused by gaps in knowledge, ambiguous prompts, or overconfident generation.

Tying model statements to verifiable sources or real-world evidence, such as retrieved documents or sensor data, to reduce unsupported claims and improve trust in generated answers.

A systematic skew in model outputs or training data that unfairly favors or disadvantages certain groups, perspectives, or outcomes, requiring measurement and mitigation throughout the lifecycle.

An attack where malicious text in user input or retrieved content manipulates a model into ignoring its instructions, leaking data, or performing unintended actions.

A privacy technique that adds calibrated noise or bounds individual contributions so that model outputs reveal little about any single training record, with a measurable privacy budget.

The degree to which humans can understand and justify a model's predictions or generated content, supported by techniques such as feature attribution, attention inspection, and counterfactual examples.