Models and training

Activation function

A function applied within a neural network layer that introduces nonlinearity or controls signal flow.

Active learning

A training approach in which a learning system selects the examples for which obtaining labels would be most useful.

Autoencoder

A model trained to encode input into a constrained representation and decode it back into a reconstruction.

Backpropagation

An efficient procedure for computing how a neural network's loss changes with respect to each parameter.

Causal language model

A model trained to predict each next token using only tokens that precede it.

Compute

Processing resources consumed by training or operating an AI system, often measured in operations, accelerator time, 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 architecture that applies learned local filters across 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 component that transforms an internal representation or prior outputs into a target output sequence or reconstruction.

Decoder-only model

A sequence model that generates tokens causally from preceding context without a separate encoder component.

Diffusion model

A generative model that learns to reverse a gradual noising process to create data from noise.

Direct preference optimization (DPO)

A preference-training method that directly adjusts a model toward preferred responses and away from rejected ones without first training a separate reward model in the classic RLHF pipeline.

Encoder

A component that transforms input into an internal representation useful for later prediction or generation.

Encoder-decoder model

A sequence model in which an encoder represents the input and a decoder generates output conditioned on that representation.

Epoch

One complete pass through the training dataset, usually divided into batches.

Federated learning

Training a shared model across distributed data holders without centralizing their raw training data.

Fine-tuning

An additional training phase that adapts a pretrained model by updating some or all parameters using task- or domain-specific data.

Generalization

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

Generative adversarial network (GAN)

A generative architecture trained through competition between a generator and a discriminator.

Gradient descent

An optimization method that iteratively changes parameters in the direction expected to reduce loss.

Graphics processing unit (GPU)

A highly parallel processor widely used to train and run neural networks.

Hyperparameter

A configuration value chosen outside ordinary parameter learning, such as learning rate, batch size, or model depth.

Instruction tuning

Fine-tuning a pretrained model on examples of instructions and desired responses so it becomes better at following task directions.

Latent space

An internal representational space whose dimensions encode learned factors or regularities in data.

Learning rate

A hyperparameter controlling the scale of parameter updates during optimization.

Loss function

A function that measures error or undesired behavior for an example or batch during model training.

Low-rank adaptation (LoRA)

A parameter-efficient fine-tuning method that freezes base-model weights and trains smaller low-rank update matrices.

Masked language model

A model trained to predict deliberately hidden tokens using context on both sides.

Mixture of experts (MoE)

A model architecture that routes each input or token through a selected subset of specialized parameter blocks rather than activating the entire model.

Model checkpoint

A saved snapshot of model parameters and related training state at a particular point.

Model distillation

Training a smaller or otherwise cheaper student model to reproduce useful behavior from a larger teacher model or ensemble.

Model weights

The learned numeric values within a model, collectively representing what training encoded into its behavior.

Objective function

A mathematical quantity a training or search process is designed to minimize or maximize.

Optimizer

The algorithm that converts gradients and training state into parameter updates.

Overfitting

When a model learns patterns specific to its training or evaluation examples and performs worse on genuinely new data.

Positional encoding

Information added to token representations so a transformer can account for order and relative position.

Pretraining

The broad initial training phase that gives a model general representations and capabilities before task-specific adaptation.

Quantization

Representing model weights or activations with lower numerical precision to reduce memory, storage, or inference cost.

Quantized low-rank adaptation (QLoRA)

A fine-tuning method that trains LoRA adapters while keeping the base model frozen in a lower-precision quantized representation.

Recurrent neural network (RNN)

A neural-network architecture that processes sequences by carrying state from one step to the next.

Regularization

A training constraint or penalty that discourages a model from fitting the training data too narrowly.

Reinforcement learning from human feedback (RLHF)

A model-alignment method that uses human preference data to train a reward signal or otherwise optimize model behavior toward preferred responses.

Training batch

A subset of training examples processed together for one optimization update or gradient estimate.

Training data

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

Transfer learning

Reusing representations or knowledge learned for one task or domain to improve another.

Variational autoencoder (VAE)

A probabilistic generative model that learns a distribution over latent representations and reconstructs or generates data by sampling from that latent space.

Frontier model

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

Open-weight model

A model whose trained parameters are available to download, inspect, or run under stated terms, without implying that the complete system is open source.

Proprietary model

A model whose weights, development artifacts, or rights to inspect, modify, run, or redistribute it remain materially controlled by an owner.

Scaling laws

Empirical relationships that estimate how model performance or loss changes as compute, data, parameters, or inference resources increase.

Small language model (SLM)

A language model deliberately kept smaller than contemporary large models to reduce resource needs or fit a narrower deployment target.

Open-source AI

An AI system made available with the artifacts and terms needed to use, study, modify, and share it, including the preferred form for making modifications.