Models and training

Activation function

A function that transforms a neural network unit's pre-activation, usually to add nonlinearity or control how much signal passes forward.

Active learning

A training loop in which the learner chooses which unlabeled examples should be sent to an oracle for labels.

Autoencoder

A model trained to encode an input into an internal representation and decode that representation into a reconstruction of the input.

Backpropagation

A reverse pass through a differentiable computation graph that computes how the loss changes with respect to each parameter.

Causal language model

A language model trained to predict each token from the tokens that precede it, without access to future positions.

Compute

The processing work or capacity used to train and run an AI system, measured with quantities such as operations, accelerator time, throughput, or cost.

Contrastive learning

A representation-learning approach that trains examples judged related to be closer and examples judged different to be farther apart in a learned space.

Convolutional neural network (CNN)

A neural network that applies shared learned filters across local regions of spatial or sequential data.

Cross-entropy

A loss that measures how poorly a predicted probability distribution represents the target distribution.

Data augmentation

Expanding or varying training examples through transformations or generation intended to preserve task-relevant meaning.

Decoder

A model component that turns an internal representation or prior outputs into a reconstruction, prediction, or target sequence.

Decoder-only model

A Transformer-style sequence model that uses causal self-attention to predict the next token without a separate encoder stack.

Diffusion model

A generative model that learns to reverse a gradual noising process, producing samples by repeatedly denoising an initial random state.

Direct preference optimization (DPO)

A preference-training method that increases a model's relative likelihood of chosen responses over rejected responses without training a separate reward model.

Encoder

A model component that turns an input into an internal representation for classification, retrieval, reconstruction, or generation.

Encoder-decoder model

A model with separate paths for representing a source input and generating a target output conditioned on that representation.

Epoch

One nominal pass through a training dataset, usually divided into batches and followed by another shuffle or sampling cycle.

Federated learning

Training a shared model from updates computed by distributed participants while their raw training data remains local.

Fine-tuning

An additional training phase that adapts a pretrained model by updating all parameters or a selected subset on new data.

Frontier model

A general-purpose AI model at or near the leading edge of broadly evaluated capability at a particular time.

Generalization

The ability of a learned model or system to perform well on relevant data, tasks, or environments that were not used to fit it.

Generative adversarial network (GAN)

A generative model trained as a game between a generator that produces candidates and a discriminator that tries to detect generated samples.

Gradient descent

An iterative optimization method that moves parameters opposite the gradient of an objective being minimized.

Graphics processing unit (GPU)

A throughput-oriented parallel processor widely used for neural-network training and inference.

Hyperparameter

A configuration choice that governs a model or learning procedure but is not ordinarily learned as a model weight.

Instruction tuning

Fine-tuning a pretrained model on instruction-and-response examples so it learns to perform tasks described in natural language.

Latent space

A space of unobserved variables or learned representations used to model structure behind observed data.

Learning rate

A hyperparameter that sets the base step size for parameter updates during optimization.

Loss function

A function that converts a model's prediction and target into a training penalty for an example or batch.

Low-rank adaptation (LoRA)

A parameter-efficient fine-tuning method that freezes selected base weights and learns low-rank matrices representing their updates.

Masked language model

A language model trained to reconstruct selected hidden tokens from the visible context on both sides.

Mixture of experts (MoE)

A model architecture in which a learned router selects one or more expert subnetworks for each input or token.

Model checkpoint

A saved snapshot of model parameters and, when needed, the surrounding state required to resume training.

Model distillation

Training a student model to reproduce useful behavior from a teacher model or ensemble, often to reduce inference cost.

Model weights

The learned numerical values in a model's connections or transformations, often used loosely for all learned parameters.

Objective function

The mathematical quantity an optimization or search process is configured to minimize or maximize.

Open-weight model

A model whose trained parameters are distributed under stated terms, without implying that its training process or complete system is open source.

Optimizer

An algorithm that turns gradients and accumulated training state into parameter updates.

Overfitting

A failure to generalize in which performance on development data is better than performance on representative unseen data.

Positional encoding

A numerical signal that lets a sequence model distinguish token positions or distances.

Pretraining

Training performed before a later adaptation stage, usually to learn reusable representations or capabilities from a broader dataset.

Proprietary model

A model whose owner materially controls access to its weights, development artifacts, operation, modification, or redistribution.

Quantization

Mapping model values to a lower-precision representation to reduce storage, memory traffic, or computation.

Quantized low-rank adaptation (QLoRA)

A fine-tuning method that backpropagates through a frozen quantized base model into trainable LoRA adapters.

Recurrent neural network (RNN)

A neural-network architecture that applies a recurrent update while carrying hidden state through a sequence.

Regularization

A change to the learning problem intended to improve performance on unseen data, often by limiting or perturbing what the model can fit.

Reinforcement learning from human feedback (RLHF)

A training approach that learns a reward signal from human judgments and uses reinforcement learning to optimize a model against that signal.

Scaling laws

Empirical relationships fitted to how model loss or performance changes with resources such as parameters, data, or compute.

Small language model (SLM)

A language model small enough for a target deployment's memory, latency, cost, or control constraints relative to contemporary large models.

Training batch

A group of training examples processed together to estimate a gradient before an optimizer update.

Training data

The examples and signals used to fit a model's learned parameters during pretraining, fine-tuning, or other learning procedures.

Transfer learning

Using representations or parameters learned in a source task or domain to help a target task or domain.

Variational autoencoder (VAE)

A latent-variable generative model trained with variational inference so it can reconstruct data and sample new examples.

Open-source AI

An AI system released with the rights and preferred form needed to use, study, modify, and share the system.