Get the 2026 ML Training Cookbook | 52 recipes — GRPO, Flow Matching, World Models, and everything in between Download Now →

Reference

Glossary

A comprehensive reference for machine learning and AI terminology — from training algorithms and model architectures to optimization, deployment, and inference techniques. Each entry includes precise definitions, key points, and related concepts.

Training

20 terms

On-Policy Distillation

A knowledge distillation technique where the student model generates its own predictions and the teacher provides corrections on those predictions, reducing train-serve distribution mismatch.

Group Relative Policy Optimization (GRPO)

A reinforcement learning algorithm that simplifies PPO by eliminating the critic network, computing advantages relative to a group of sampled outputs per prompt.

LoRA (Low-Rank Adaptation)

A parameter-efficient fine-tuning method that freezes the pre-trained model and injects trainable low-rank decomposition matrices into each layer.

Fine-Tuning

The process of adapting a pre-trained model to a specific task or domain by continuing training on task-specific data.

Knowledge Distillation

A model compression technique where a smaller student model learns to mimic the behavior of a larger, more capable teacher model.

RLHF (Reinforcement Learning from Human Feedback)

A technique that aligns language models with human preferences by training a reward model on human feedback and optimizing the policy via reinforcement learning.

DPO (Direct Preference Optimization)

A preference optimization method that directly optimizes the policy from preference data without training a separate reward model.

PPO (Proximal Policy Optimization)

A policy gradient RL algorithm that uses clipped surrogate objectives to ensure stable, conservative policy updates.

PEFT (Parameter-Efficient Fine-Tuning)

A family of techniques that fine-tunes a small subset of model parameters while freezing the rest, drastically reducing compute and memory requirements.

Supervised Fine-Tuning (SFT)

Fine-tuning a pre-trained model on labeled input-output pairs for a specific task, forming the first stage of alignment pipelines like RLHF.

Instruction Tuning

A form of fine-tuning that trains models to follow natural language instructions across diverse tasks, improving generalization and zero-shot capability.

Batch Size

The number of training examples processed in a single forward and backward pass, directly affecting memory usage, gradient quality, and training speed.

Learning Rate

A hyperparameter controlling the step size of weight updates during training, critically affecting convergence speed and final model quality.

Gradient Accumulation

A technique that simulates larger batch sizes by accumulating gradients across multiple forward passes before applying a weight update.

Regularization

A set of techniques that constrain or penalize model complexity to prevent overfitting and improve generalization to unseen data.

Dropout

A regularization technique that randomly deactivates a fraction of neurons during training, preventing co-adaptation and improving generalization.

Cross-Entropy Loss

A loss function measuring the difference between predicted probability distributions and true labels, standard for classification and language modeling.

Backpropagation

The algorithm used to compute gradients of the loss with respect to all model weights by applying the chain rule backward through the network.

Early Stopping

A regularization technique that halts training when validation performance stops improving, preventing overfitting.

Weight Decay

An L2 regularization technique that penalizes large weights by adding their squared magnitude to the loss, encouraging simpler models.

Architecture

9 terms

Flow Matching

A generative modeling framework that learns a vector field transporting a simple prior to the target data distribution, offering more flexibility and faster sampling than diffusion.

World Models

AI systems that learn internal representations of environment dynamics, enabling prediction and planning in latent space for dramatically improved sample efficiency.

Small Language Model (SLM)

A language model with significantly fewer parameters than frontier models, optimized for efficiency, cost, and on-device deployment.

Tokenization

The process of breaking text into discrete units (tokens) that a model can process, bridging raw text and numerical representations.

Attention Mechanism

A neural network component that computes weighted relationships between all positions in a sequence, enabling models to focus on relevant context.

Transformer Architecture

A neural network architecture based entirely on self-attention, replacing recurrence with parallel attention layers for superior sequence modeling.

Mixture of Experts (MoE)

An architecture that routes each input to a subset of specialized expert networks, increasing model capacity without proportional compute increase.

Embedding

A learned dense vector representation of discrete tokens or entities that captures semantic relationships in continuous space.

Latent Space

A learned lower-dimensional continuous representation that captures the essential structure of data, enabling efficient manipulation and generation.

Optimization

3 terms

Quantization

A technique that reduces the precision of model weights and activations (e.g., from FP32 to INT8) to decrease memory usage and accelerate inference.

Mixed Precision Training

A training technique that uses lower-precision formats (FP16/BF16) for computations while maintaining FP32 master weights for numerical stability.

Hyperparameter Tuning

The process of finding optimal hyperparameter values (learning rate, batch size, etc.) that maximize model performance on validation data.

Deployment

1 term

ONNX (Open Neural Network Exchange)

An open format for representing machine learning models that enables interoperability between training and inference frameworks.

Inference

4 terms

Evaluation