Introduction to AI, Large Language Models, GenAI

Flashcards based on the concepts from the lecture on Artificial Intelligence and Generative AI.

Generative AI

A subset of AI that can create new content such as text, images, videos, and audio.

Large Language Models (LLMs)

Models trained on massive text datasets using self-supervised learning that can understand and generate human-like text.

Transfer Learning

A technique that leverages pre-trained models to enhance and expedite the development of new models.

Prompt Engineering

The process of crafting effective prompts to guide model outputs towards relevance and accuracy.

Generative Adversarial Networks (GANs)

A type of generative model architecture used for creating realistic content by pitting two neural networks against each other.

Variational Autoencoders (VAEs)

A generative model used to create new data instances, primarily in image generation and anomaly detection.

Fine-tuning

The process of adapting a pre-trained model to a specific task by continuing the training on a new dataset.

Neural Networks

Computational models inspired by the human brain that consist of interconnected nodes (neurons) for processing data.

Backpropagation

A training algorithm for neural networks that adjusts weights by propagating the error back through the network.

Reinforcement Learning with Human Feedback (RLHF)

A technique used to align AI models with human values and ethics by incorporating human feedback into the training process.

OpenAI GPT (GPT-3.5, GPT-4)

Models that excel with clear instructions and contextual examples, using techniques like few-shot prompting, Chain-of-Thought (CoT) prompting, role-based instructions, iterative refinement, and system prompts.

Few-shot prompting

A technique that improves task-specific performance by providing examples within the prompt.

Chain-of-Thought (CoT) prompting

A prompting method that enhances reasoning for complex tasks by breaking down problems into a series of steps.

Role-based instructions

Guidance in prompts that specifies the role of the model, such as 'You are a data scientist,' to influence behavior.

Iterative refinement

A process that ensures precision by allowing the model to revise its outputs.

System prompts

a specialized type of prompt used in AI language models that sets the overall context, behavior, or persona for the AI's responses.

EX: Do not create images in the style of artists, creative professionals or studios whose latest work was created after 1912 (e.g. Picasso, Kahlo).

They guide the model's output by defining constraints and expectations.

Google PaLM (PaLM 2)

Models that benefit from CoT for reasoning, few-shot prompting for accuracy, and generated knowledge prompting for multi-step queries.

Meta LLaMA (LLaMA 2, LLaMA 3)

Models that excel in in-context learning, structured dialogue prompts, text-to-SQL formatting, and prompt chaining.

Anthropic Claude (Claude 2, Claude 3)

Models that respond well to XML-tagged prompts, utilize step-by-step reasoning (CoT), assign roles, and leverage long context for handling extensive inputs.

Code LLaMA

Models that excel at few-shot examples for code generation, function calling prompts, and debugging workflows.

LoRA (Low-Rank Adaptation)

A finetuning framework that modifies only a fraction of a model’s parameters, using low-rank matrices to adjust large models efficiently without retraining the entire architecture.

QLoRA (Quantized Low-Rank Adaptation)

A finetuning approach that combines LoRA with quantized precision to optimize memory and computation overhead while fine-tuning large language models.

Parameter-Efficient Fine-Tuning (PEFT)

Techniques that fine-tune a pretrained model by updating only a small subset of its parameters, offering improved computational efficiency.

Supervised Fine-Tuning (SFT)

The process of adapting a pretrained model on labeled data to align its responses with specific tasks or instructions.

Preference Tuning

A finetuning process aimed at aligning an LLM's behavior with desired human preferences through training based on a reward model.

Direct Preference Optimization (DPO)

A technique in preference tuning that directly compares outputs from different models to optimize for preferred generations during training.

Proximal Policy Optimization (PPO)

A more complex algorithm compared to DPO used in training LLMs to align outputs with preferences.

Benchmarking

The process of evaluating and comparing models across various metrics like accuracy, latency, and memory efficiency to assess finetuning success.

Metrics for evaluation (BLEU, F1-score, perplexity)

Quantitative measures used to evaluate the performance of generative models, providing insight on accuracy and quality.

Human evaluation

Methods of assessing model outputs through human judgment, often providing the most comprehensive insight but being resource-intensive.

ROUGE

A metric used for evaluating text generation that compares the generated text to reference text, assessing recall and precision at the n-gram level.

MMLU (Massive Multitask Language Understanding)

A benchmark that assesses the proficiency of language models across a variety of tasks and domains.

GLUE (General Language Understanding Evaluation)

A benchmark that evaluates and measures the performance of language models across multiple language understanding tasks.

Chatbot Arena

A human evaluation method that uses structured comparisons to assess the quality and performance of conversational agents.

Retrieval-Augmented Generation (RAG)

A framework that enhances language models by integrating external information retrieval to improve the accuracy and relevance of generated content.

Semantic search

A search method that finds contextually relevant information by encoding user queries and documents into vector representations, identifying the most conceptually similar documents.

Vector databases

Databases that store vector embeddings, allowing for efficient similarity searches and real-time updates as new information becomes available.

Indexing

The process of processing and structuring external data sources, splitting documents into smaller chunks, and converting these into embeddings for efficient retrieval.

Embedding

A vector representation of text or other data that captures its semantic meaning, allowing for similarity comparisons during retrieval.

Dense Retrieval

A retrieval method that maps text segments into a dense vector space of features for efficient searching.

Sparse Retrieval

A retrieval method that indexes text segments using specific words for exact matching.

Chunking

The process of dividing documents into smaller, more manageable segments for indexing and retrieval.

Query Rewriting

The process of modifying a user’s query to improve accuracy and relevance in search results.

Hallucination

The tendency of language models to generate false or nonsensical information that is not grounded in reality.

Chunking Optimization

The practice of experimenting with different chunk sizes and strategies to balance semantic coherence and noise reduction in retrieval tasks.

Knowledge Graph

A graph-structured database that represents entities and their relationships, facilitating dynamic retrieval of knowledge.

Data Dependency

The subset of data segments indispensable for addressing a query.

In-domain Data

Data from the same domain as the task at hand, improving the model's relevance and accuracy.

Zero-Shot Learning

A type of machine learning where a model can perform a task without having seen any specific examples of that task during training.

Few-Shot Learning

A type of machine learning where a model can learn a new task from only a few examples.

In-context Learning

The ability of a language model to learn from examples provided in the prompt, without requiring explicit fine-tuning.

IR (Information Retrieval)

The process of obtaining information system resources that are relevant to an information need from a collection.

Breadth-First Search (BFS)

An algorithm for traversing or searching tree or graph data structures.

Fine-tuning

The process of further training a pre-trained language model on a smaller, domain-specific dataset.

Instruction Tuning

Supervised fine-tuning using paired (instruction, output) data to infuse new capabilities into LLMs.

G-Evals

NLG evaluation using GPT-4 with better human alignment for offline evaluation techniques

Agent

A system that leverages a language model to determine which actions it should take and in what order, often using external tools.

Chains

A LangChain concept for connecting LLMs with additional components like prompt templates, external tools, or other LLMs, extending their capabilities.

ConversationBufferMemory

A type of memory in LangChain that stores the entire conversation history for an LLM to reference.

ConversationBufferWindowMemory

A type of memory in LangChain that only retains the last k number of conversations in the history.

ConversationSummaryMemory

A type of memory in LangChain that summarizes the entire conversation history to distill it into the main points, using another LLM for summarization.

GGUF

A file format for storing quantized models, often used with llama.cpp.

LangChain

A framework designed to simplify working with LLMs through useful abstractions and modular components.

LLMChain

A specific type of chain in LangChain that combines an LLM with a prompt and memory to streamline interactions.

Model I/O

The process of loading and working with LLMs, including handling input and output formats.

Open LLM Leaderboard

A ranking of open-source LLMs based on performance metrics.

Prompt Template

A predefined structure for formatting input prompts to an LLM, ensuring consistency and reusability.

Quantization

A technique for reducing the number of bits required to represent the parameters of an LLM, resulting in faster run times and reduced memory usage.

ReAct (Reasoning and Acting)

A framework for prompting LLMs that combines reasoning and acting in an iterative process of thought, action, and observation.

Sequential Chains

A series of connected chains in LangChain where the output of one chain is used as the input for the next, allowing for complex tasks to be broken down into smaller subtasks

Knowledge Graph

A powerful data structure that represents information as a network of interconnected entities and relationships, mimicking human-like understanding of concepts.

Nodes

Elements in a knowledge graph that represent entities.

Edges

Connections in a knowledge graph that represent relationships between entities.

Entity

An object or concept within a knowledge graph, represented by a node.

Relationship

The connection between entities in a knowledge graph, represented by an edge.

LLMs and Knowledge Graphs

Serve as a bridge between unstructured data and machine-readable formats, enhancing the reasoning capabilities of language models.

Multimodal Models

Models designed to process and integrate different types of data, such as text, images, and audio.

Vision-Enhanced Large Language Models (Vision LLMs)

Models that combine visual and textual information for tasks like image captioning and visual question answering.

Cross-Attention Mechanisms

Techniques used in vision LLMs to align visual features with textual representations.

Diffusion Models

Models primarily used for image generation that learn to reverse a noise-adding process to create realistic images from noise.

Image-to-Text Generation

The adaptation of diffusion models for generating text based on image content.

Dataset Preparation for Vision LLMs

Focuses on aligning visual and textual features, emphasizing data cleaning, normalization, and tokenization.

Dataset Preparation for Diffusion Models

Concentrates on image normalization and augmentation to enhance visual diversity.

Image Quality and Variety

Factors that diffusion models are particularly sensitive to, which affect the generation of realistic outputs.

Data Augmentation

Techniques used to artificially expand the size of a dataset by creating modified versions of existing data.