LLM Terms

Question 1

Large Language Model (LLM)

Answer 1

A type of neural network model designed to understand and generate human-like text by learning from vast amounts of language data.

Question 2

Transformer Architecture

Answer 2

A deep learning model architecture that uses self-attention mechanisms to process input data in parallel, making it the foundation of modern LLMs like GPT and BERT.

Question 3

Self-Attention

Answer 3

A mechanism that enables the model to focus on different parts of the input sequence, assigning importance scores to each token relative to others in the context.

Question 4

Context Length

Answer 4

The maximum number of tokens an LLM can process in a single input. Longer context lengths allow the model to capture more extensive dependencies in the input text.

Question 5

Tokenization

Answer 5

The process of converting text into smaller units (tokens) that the LLM can process, such as words, subwords, or characters.

Question 6

Subword Tokenization

Answer 6

A method of breaking words into smaller units (subwords), enabling the model to handle out-of-vocabulary words and morphological variations.

Question 7

Byte-Pair Encoding (BPE)

Answer 7

A tokenization technique that iteratively merges the most frequent pairs of characters or character sequences to create subword tokens.

Question 8

Prompt Engineering

Answer 8

The practice of crafting input prompts to guide the LLM in generating specific, desired outputs, essential for steering model behavior in applications.

Question 9

Few-Shot Learning

Answer 9

A model’s ability to learn and generalize from only a few examples provided in the prompt, enabling the generation of contextually relevant responses.

Question 10

Zero-Shot Learning

Answer 10

The capability of an LLM to perform a task without explicit examples in the prompt, relying instead on its pre-trained knowledge.

Question 11

Fine-Tuning

Answer 11

The process of training a pre-trained LLM on a smaller, task-specific dataset to adapt the model for specific applications.

Question 12

Retrieval-Augmented Generation (RAG)

Answer 12

A technique combining information retrieval and LLMs, where relevant documents are retrieved and provided as context to the model during inference.

Question 13

Knowledge Base

Answer 13

A structured repository of information used to provide factual data for an LLM, enhancing its ability to generate accurate and contextually appropriate responses.

Question 14

Semantic Search

Answer 14

A search method that leverages the meaning of words and their relationships to find relevant documents or information, often used in conjunction with LLMs.

Question 15

Embeddings

Answer 15

Dense vector representations of words, sentences, or documents that capture their semantic meanings and are used for tasks like semantic search in LLM applications.

Question 16

Memory Management

Answer 16

Techniques used in LLMs to handle the model’s context length, including how to store and retrieve information across long conversations.

Question 17

Temperature

Answer 17

A parameter that controls the randomness of an LLM’s output. Higher values generate more diverse outputs, while lower values make responses more deterministic.

Question 18

Top-k Sampling

Answer 18

A decoding method that restricts the model’s predictions to the top k most likely tokens at each step, enhancing the quality of generated text.

Question 19

Top-p (Nucleus) Sampling

Answer 19

A sampling method that dynamically selects tokens based on their cumulative probability distribution, allowing for more flexible and coherent text generation.

Question 20

Beam Search

Answer 20

A decoding strategy that explores multiple possible sequences during text generation, keeping the top n most probable sequences at each step.

Question 21

Hallucination

Answer 21

A phenomenon where an LLM generates content that is not factually accurate or related to the given context, often seen as a critical challenge in LLM applications.

Question 22

Latent Space

Answer 22

The high-dimensional space in which LLMs represent input data, capturing complex relationships and semantics between words and sentences.

Question 23

Embedding Models

Answer 23

Models trained to generate vector representations of input data (words, sentences, documents), often used in combination with LLMs for tasks like semantic search.

Question 24

Sequence-to-Sequence (Seq2Seq) Model

Answer 24

A model architecture that maps input sequences to output sequences, commonly used in tasks like translation and summarization.

Question 25

Attention Mask

Answer 25

A mechanism that controls which parts of the input sequence an LLM should focus on, used to handle padding and manage context in sequence processing.

Question 26

Positional Encoding

Answer 26

A technique used in Transformer models to provide information about the order of tokens in the input sequence.

Question 27

Encoder-Decoder Architecture

Answer 27

A neural network framework that consists of an encoder processing the input sequence and a decoder generating the output sequence, foundational to many NLP tasks.

Question 28

Pre-training

Answer 28

The initial phase of training an LLM on a large corpus of text data to learn general language patterns and representations.

Question 29

Transfer Learning

Answer 29

Adapting a pre-trained LLM to a new task or domain by fine-tuning it on a smaller, specialized dataset.

Question 30

Language Model Inference

Answer 30

The process of generating output text from an LLM given an input prompt or query.

Question 31

Hyperparameter Tuning

Answer 31

Adjusting model parameters (e.g., learning rate, batch size, temperature) to optimize the performance of LLMs for specific tasks.

Question 32

Model Checkpointing

Answer 32

Saving intermediate states of an LLM during training, allowing for recovery and further fine-tuning if necessary.

Question 33

Gradient Accumulation

Answer 33

A technique used to overcome memory limitations during LLM training by accumulating gradients over multiple batches before updating model weights.

Question 34

Distributed Training

Answer 34

The process of training an LLM across multiple GPUs or machines to handle large datasets and model sizes efficiently.

Question 35

Data Augmentation

Answer 35

Techniques for increasing the diversity of training data, such as paraphrasing or adding noise, to improve LLM performance.

Question 36

Text Summarization

Answer 36

The task of generating a concise and coherent summary of a longer text, often facilitated by LLMs in various applications.

Question 37

Named Entity Recognition (NER)

Answer 37

A technique for identifying and classifying entities (e.g., names, dates, locations) in text, used to extract structured information in LLM applications.

Question 38

Conversational Agent

Answer 38

A system that uses an LLM to interact with users in natural language, often designed to handle tasks like customer support, information retrieval, or personal assistance.

Question 39

Conversational Memory

Answer 39

Mechanisms in an LLM application that retain context across multiple user interactions to facilitate coherent dialogue.

Question 40

Latency

Answer 40

The time delay between a user input and the model’s response, a critical factor in the user experience of LLM-based applications.

Question 41

Inference Optimization

Answer 41

Techniques such as model quantization or pruning to reduce the computational cost and latency of LLMs during inference.

Question 42

Knowledge Distillation

Answer 42

The process of transferring knowledge from a large, complex LLM to a smaller, more efficient model to improve performance in resource-constrained environments.

Question 43

Prompt Tuning

Answer 43

The process of adjusting prompts used with an LLM to elicit more accurate, relevant, or contextually appropriate responses.

Question 44

Bias in Language Models

Answer 44

The presence of stereotypes or unfair associations learned from training data, necessitating strategies to mitigate bias in LLM applications.

Question 45

Zero-Shot Transfer

Answer 45

The ability of an LLM to perform tasks in new domains without explicit retraining on domain-specific data.

Question 46

Conversational Turn

Answer 46

A single interaction between a user and an LLM in a dialogue, which includes both the user input and the model’s response.

Question 47

Contextual Embeddings

Answer 47

Word embeddings that change depending on the context in which a word appears, enabling more nuanced understanding in LLMs.

Question 48

Query Parsing

Answer 48

The process of analyzing and interpreting a user’s input to understand intent and extract relevant information for the LLM application.

Question 49

Memory Augmented LLM

Answer 49

An LLM equipped with external memory components to store and retrieve information beyond its fixed parameters, enhancing its capability to answer specific queries.

Question 50

Scalability

Answer 50

The ability of an LLM application to handle an increasing amount of work, data, or users without sacrificing performance or accuracy.