liveBookmanning.com

What you will learn in Generative AI in Action

by Amit Bahree

Read it on liveBook Get the book at manning.com

Chapter by chapter

Concepts explained in this book

• language modelA large language model is a neural model trained on large text corpora to understand and generate natural language, including source code in code-focused variants.
• fine-tuningTask-specific fine-tuning is the adaptation of a pretrained model using data targeted at a particular task.
• promptA system prompt is an instruction supplied to a generative model that establishes behavior, information, or constraints for its responses.
• vector databaseA vector database is a data-management system that stores and provides access to data together with numerical vector representations and supports vector-based retrieval.
• agent orchestrationContainer orchestration is the automated deployment, scheduling, scaling, networking, and management of containers.
• generative modelingGenerative AI is an artificial-intelligence approach that produces new content or processes rather than merely classifying or predicting properties of existing inputs.
• prompt engineeringPrompt engineering is the deliberate design and refinement of inputs provided to generative models to elicit desired outputs.
• evaluationEvaluation is the systematic assessment of a model's generated outputs against specified criteria or requirements.
• foundational modelA pretrained model is a machine-learning model that has already been trained on an initial dataset before being adapted to another task or dataset.
• similarity searchRetrieval is the process of selecting relevant information from a collection in response to an information need.
• benchmarkBenchmarking is the systematic measurement and comparison of models or systems against defined tasks, datasets, or criteria.
• retrieval-augmented generationRetrieval-augmented generation is a generation technique that retrieves relevant external information and supplies it to a generative model as context for its response.
• generationNatural language generation is the computational production of human-readable language from structured or unstructured inputs.
• human-in-the-loopHuman evaluation is the assessment of an artifact or system by people using judgments, observations, surveys, or interviews.
• conversational aiA chatbot is software that conducts conversational interactions with users through text or speech.
• performancePerformance is the degree to which a system achieves operational objectives such as response speed, throughput, and resource efficiency.
• context windowThe context window is the bounded amount of input and generated text that a language model can consider within one interaction.
• text chunkChunking is the partitioning of data, especially text, into smaller units for processing, indexing, or retrieval.
• small language modelA small language model is a language model with fewer parameters and lower computational, training, and operational requirements than larger language models.
• scalabilityScalability is the ability of a system to accommodate increasing workloads, users, or traffic while maintaining acceptable operation.
• code generationCode generation is the computational production of source code by a model from learned programming-language patterns and inputs.
• training datasetTraining data is the collection of examples or information from which a machine-learning model learns its patterns or parameters.
• monitoringLogging is the recording of software events, states, and diagnostic information for operational analysis and troubleshooting.
• tokenA token is a discrete unit into which text or other model input is encoded for processing by a language model.
• chat APIA chat completion API is a programming interface that generates assistant responses from a sequence of conversational messages.
• red-teamingA red team is an authorized group that deliberately probes a system for vulnerabilities, failure modes, or harmful behavior.
• adaptabilityModel adaptation is the adjustment of a generative AI model so that its outputs perform better for particular tasks or domains.
• responsible AIResponsible AI is the practice of developing and operating artificial intelligence in ways that address safety, fairness, accountability, transparency, and related societal risks.
• authorizationAccess control is the practice of restricting access to systems, resources, or data according to identity, policy, or authorization.
• prompt injectionJailbreaking is an attack technique that manipulates an AI model into bypassing intended instructions, safeguards, or behavioral constraints.
• FAISSFAISS is a library for efficient similarity search and clustering of dense vectors on CPUs and GPUs.
• RedisRedis is an open-source in-memory data store that supports key-value storage and multiple data structures for database, caching, and messaging workloads.
• data governanceData governance is the framework of policies, responsibilities, and controls for managing data quality, access, use, and compliance.
• migrationETL is a data-processing pattern that extracts data from sources, transforms it, and loads it into a destination system.
• evaluation datasetA validation set is a dataset reserved for assessing or guiding model performance during development rather than fitting the model parameters.
• transparencyTransparency is the degree to which an AI system's operation, information sources, limitations, and decision processes are made visible to relevant people.
• throughputThroughput is the amount of work or number of requests a system processes during a specified period.
• vector indexA vector index is a data structure that organizes high-dimensional vectors to support efficient similarity and nearest-neighbor searches.
• encoderA text encoder is a neural-network component that converts textual input into an internal vector representation.
• Chroma DBChroma is an open-source vector database for storing and querying embeddings and their associated documents or metadata.
• LangChainLangChain is a software framework that provides modular components for building applications around large language models and related natural-language services.
• LlamaIndexLlamaIndex is a data framework that connects large language models with private data sources through ingestion, indexing, retrieval, and query interfaces.
• semantic functionA semantic function is a Semantic Kernel callable operation defined by a natural-language prompt and input parameters.
• Semantic KernelSK is an AI application framework that provides abstractions for integrating language models with prompts, functions, and external services.
• sensitive informationEnterprise data is organization-owned information that may be subject to access restrictions, privacy requirements, and legal obligations.
• reinforcement learning with human feedbackReinforcement learning from human feedback is a model-training technique that uses human preferences as feedback to optimize generated behavior.
• Azure OpenAIAzure OpenAI on your data is a platform service that enables enterprises to run AI models over their organizational data with integrated scalability, security, and Azure-service connectivity.
• embeddingA vector is an ordered sequence of numerical values representing a point or quantity in a multidimensional space.
• completions apiThe completion API is an interface that generates text completions from a supplied input sequence and can return them incrementally with token-level metadata.
• LLMOpsLLMOps is the set of operational practices for developing, deploying, monitoring, governing, and maintaining large language models and applications built with them.
• cachingCaching is the practice of storing reusable data or results in faster-access storage so they can be returned without repeating the original computation or request.
• token limitA token limit is a maximum number of tokens that a model may process or generate in an operation.
• groundingGrounding is the technique of supplying a generative model with relevant external information as context for its response.
• observabilityObservability is the practice of monitoring, logging, and tracing a system to understand its internal behavior and diagnose problems from its outputs and operational data.
• content filteringContent moderation is the screening, classification, and handling of content to identify or restrict material that violates safety, policy, or appropriateness requirements.
• continuous improvementA development method that repeatedly evaluates outcomes and refines a system through successive cycles.
• variational autoencoderA variational autoencoder is a generative neural model that encodes data into a probabilistic latent representation and decodes samples from that representation into reconstructed or newly generated data.
• hallucinationAn AI hallucination is generated content that is coherent or plausible but factually unsupported, nonsensical, or fabricated.
• cosine similarityCosine similarity measures the cosine of the angle between two vectors and indicates their directional similarity.
• rate limitThrottling is the deliberate slowing or limiting of requests or operations to control resource use and system load.
• low-rank adaptationLow-rank adaptation is a parameter-efficient technique that adapts a pretrained model by training small low-rank matrices while leaving the original model weights frozen.
• Redis SearchRedis Search is a Redis module that provides text search, secondary indexing, and vector similarity search capabilities.
• temperatureTemperature is a model-generation setting that controls the degree of randomness in the model's output distribution.
• baselineA baseline model is a reference model whose performance provides a comparison point for evaluating another model.
• few-shot learningFew-shot learning is the adaptation of a model to a task by providing a small number of task-specific examples at use time.
• G-EvalG-Eval is an LLM-based framework for evaluating generated text according to specified quality criteria without requiring reference texts.
• fairnessFairness is the property of treating people equitably and avoiding unjustified disparities in the outcomes or treatment produced by an AI system.
• preprocessingData preprocessing is the process of cleaning, transforming, and organizing raw data into a form suitable for analysis or model processing.
• validationValidation is the process of checking generated software or content against functional, quality, security, and compliance requirements.
• accuracyAccuracy is the proportion or degree of evaluated outputs that are correct according to a specified reference or criterion.
• f-scoreF-measure is a metric that combines precision and recall into a single balanced score.
• biasBias is a systematic tendency in data or model behavior that produces unfair, distorted, or unbalanced outcomes.
• overfittingOverfitting is a modeling condition in which learned patterns fit training data too closely and fail to generalize well to new examples.
• generative ai application stackA generative AI application stack is a layered architecture of models, services, infrastructure, interfaces, and operational components that supports generative AI applications.
• GPT modelA family of generative pretrained transformer language models developed by OpenAI.
• reliabilityReliability is the ability of a system to perform its required functions consistently and correctly over time and under expected conditions.
• hybrid searchHybrid search combines multiple search methods, such as text-based, categorical, numerical, and vector-based retrieval, in one query.
• privacyPrivacy is the protection of personal or confidential information from inappropriate collection, access, use, or disclosure.
• transfer learningTransfer learning adapts knowledge learned from one task or dataset to improve performance on another task.
• diffusion modelA diffusion model is a generative model that learns to create data by reversing a process that progressively adds noise to examples.
• trajectoryTracing is the technique of following a request through the operations and services involved in processing it.
• ROUGEROUGE is a family of automatic metrics that evaluates generated text by measuring overlap with reference text.
• redundancyA redundancy strategy provides duplicate or alternate system resources so service can continue when a component fails or becomes unavailable.
• uptimeThe property of a system remaining operational and accessible for a high proportion of time.
• neural network layerA neural-network layer is a compositional processing stage that transforms representations through connected computational units and learned parameters.
• open-source modelAn open-source large language model is a language model whose source artifacts and permitted use are made available under an open-source license.
• load balancingLoad balancing distributes computational requests or workloads across multiple resources to improve utilization, performance, and availability.
• inpaintingInpainting is the generation or replacement of selected regions within an existing image while preserving or incorporating surrounding visual context.
• self-consistencyA self-consistency check compares multiple generated results or reasoning paths to identify inconsistent or unreliable outputs.
• quantizationQuantization is the reduction of numerical precision used to represent model parameters, typically to reduce memory use and computational cost.
• coherenceCoherence is the degree to which generated content is logically consistent and meaningfully connected.
• fluencyLanguage fluency is the grammaticality, naturalness, and readability of generated language.
• indexingData indexing is the organization of stored data into searchable structures that support efficient information retrieval.
• latent spaceA latent space is a lower-dimensional representation space in which data items are encoded by features that capture meaningful structure and similarity.
• relevanceAnswer relevancy is a metric that measures how directly a generated answer addresses the question or information need.
• video generationA video generator is a generative system that produces video content from user-specified inputs such as prompts.
• BLEU scoreBLEU is an automatic metric that evaluates machine-translated text by measuring its n-gram overlap with a human reference translation.
• chunkingFixed-length chunking divides text into chunks according to a predetermined character, word, or token length rather than linguistic boundaries.
• Vision TransformerA vision transformer is a Transformer-based neural model that processes images by converting image regions into a sequence of representations and applying attention.
• generalizationGeneralization is a model's ability to perform effectively on examples that were not present in its training data.
• underfittingUnderfitting is a modeling condition in which a model fails to learn enough structure from its training data to perform well on new examples.
• generative adversarial networkA generative adversarial network is a generative model composed of a generator and discriminator trained adversarially to produce realistic synthetic data.
• promptopsPrompt operations is the systematic management, evaluation, deployment, and maintenance of prompts and prompt-based workflows.
• prompt flowPrompt Flow is a Microsoft open-source toolset for designing, evaluating, testing, and deploying workflows that combine prompts, language models, and code.
• response filteringA response filtering system evaluates and suppresses or modifies generated outputs that violate safety, content, or compliance requirements.
• complianceCompliance is conformity with applicable laws, regulations, standards, policies, and contractual requirements.
• classifierText classification is the assignment of text to one or more predefined categories.
• sentence splittingSentence splitting, also called sentence segmentation, identifies sentence boundaries to divide text into sentence-level units.
• commercial modelA generative AI model distributed or operated under proprietary commercial terms rather than an open-source license.
• image generationImage generation is the computational production of visual images from learned patterns or specified inputs.
• multimodalA multimodal model is a model that processes or generates information in multiple modalities such as text, images, audio, or video.
• tokenizationTokenization is the process of converting text into a sequence of discrete tokens for language-model processing.
• fidelityGroundedness is the degree to which an AI output conforms to the information in its sources rather than speculation or fabrication.
• role-based access controlRole-based access control is an authorization method that assigns permissions to roles and grants users access through those roles.
• low-rank matrixA low-rank matrix is a matrix whose rank is substantially smaller than its number of rows or columns, allowing a compact representation of a transformation.
• MLflowMLflow is an open-source platform for managing machine-learning experimentation, reproducibility, model packaging, deployment, and lifecycle operations.
• OpenLLMetryOpenLLMetry is an open-source observability extension that provides instrumentation and telemetry for large language model operations.
• TraceloopTraceloop is an observability tool and service for monitoring, alerting, and tracing large language model applications.
• normalizationNormalization is the transformation of data into a consistent scale, representation, or format.
• HNSWHNSW is a graph-based approximate nearest-neighbor indexing algorithm that organizes vectors in navigable hierarchical layers.
• corpusA corpus is a collection of documents or passages assembled for information retrieval or language processing.
• securitySecurity is the protection of systems, data, and services against unauthorized access, alteration, disruption, or destruction.
• objective functionTraining loss is a metric measuring the error between a model’s predictions and actual outcomes on training examples.
• style transferStyle transfer is the transformation of an image so that stylistic characteristics from one source are applied to another image while preserving its content.
• retrieval indexA document index is a data structure that organizes documents or their representations to support efficient information retrieval.
• zero-shot learningZero-shot learning is the ability to perform a task or recognize classes not directly represented in task-specific training examples by relying on broader learned associations or descriptions.
• precisionPrecision is the proportion of items identified as positive by a system that are actually positive.
• recallRecall is the proportion of relevant or actual positive cases that a system correctly identifies.
• cost optimizationCost management is the planning, monitoring, and control of resources and expenditures to achieve objectives within financial constraints.
• Azure Content SafetyAzure Content Safety is a Microsoft Azure service for analyzing and moderating potentially harmful, ungrounded, or protected content in AI inputs and outputs.
• prompt shieldA prompt shield is a content-safety mechanism that identifies and neutralizes direct or indirect prompt attacks by distinguishing trusted instructions from untrusted input.
• protected material detectionProtected material detection is the identification of generated content that matches copyrighted or otherwise owned source material.
• SpotlightingSpotlighting is an ensemble of techniques for helping language models distinguish valid system instructions from potentially untrustworthy external input.
• attentionAttention is a neural-network mechanism that assigns different importance weights to parts of an input when producing an output.
• generatorA text generator is a model or model component that produces natural-language text from input context.
• schemaA schema is a formal specification of the structure and permitted content of a data representation.
• SoraSora is an AI model that generates video scenes from textual descriptions.
• text-to-video generationText-to-video generation is the synthesis of video content from textual descriptions.
• narrow aiTraditional AI is the class of artificial-intelligence systems primarily characterized here as making predictions from input data rather than generating novel content.
• time per output tokenTime per output token is the average time required to generate each output token, typically excluding the time attributable to processing the input prompt.
• time to first tokenFirst token latency is the elapsed time from a request until the first generated token is returned by a generative model.
• sanitizationSanitization is the transformation or removal of unsafe content from input or output before it is processed or rendered.
• chain-of-thought promptingChain of thought is a prompting technique that decomposes a problem into intermediate reasoning steps before producing an answer.
• epochAn epoch is one complete pass through the examples in a training dataset during model training.
• emergent behaviorAn emergent property is a capability or behavior that arises from a system's components or scale without being explicitly programmed as an individual feature.
• nondeterminismNondeterminism is the property that repeated executions with the same apparent inputs can produce different outputs rather than a single fixed result.
• adversarial inputAn adversarial example is an input modified or crafted to cause a machine-learning model to produce an incorrect or undesirable output.
• Phi-2Phi-2 is a 2.7-billion-parameter small language model developed by Microsoft.
• tokens per minuteTPM is a throughput metric expressing the number of tokens processed per minute by a language-model service.
• decodeA decoder is a neural-network component that generates or reconstructs output data from an internal representation or sampled latent state.
• Stable DiffusionStable Diffusion is a family of diffusion-based generative models that produces images from textual and other conditioning inputs.
• design for failureA systems-engineering pattern that anticipates component failures and provides fallback, recovery, or degradation behavior.
• NLTKNLTK is an open-source Python library providing interfaces and tools for natural-language processing, corpora, lexical resources, and text processing.
• PunktSentenceTokenizerPunktSentenceTokenizer is an unsupervised statistical tokenizer that identifies sentence boundaries in text.
• nlp-based chunkingNLP-based chunking is the division of text into units according to linguistic structure or meaning, such as clauses, phrases, or entities.
• spaCyspaCy is an open-source Python library for efficient natural-language processing, including linguistic annotation and text segmentation.
• mean token accuracyToken accuracy is the proportion of tokens predicted correctly in a model output sequence.
• validation lossValidation loss is a metric measuring a model’s prediction error on examples held out for validation.
• vectorizationVectorization is the transformation of data into numerical vectors suitable for computation, comparison, or model input.
• GeminiGemini is Google's multimodal AI model family for understanding and generating text, images, video, audio, and code.
• flatFLAT is a brute-force vector search approach that compares a query with every indexed vector.
• mlopsMLOps is the set of practices for developing, deploying, monitoring, and maintaining machine-learning models in operational environments.
• task-specific datasetA task-specific dataset is a collection of examples selected and prepared for training or evaluating a model on a particular task.
• QLoRAQLoRA is a low-rank adaptation technique that quantizes adapter parameters to reduce the memory required for fine-tuning large language models.
• log probabilityLogprobs are completion metadata containing the natural logarithms of the model probabilities assigned to candidate tokens.
• stress testStress testing is the evaluation of a system under demanding, unusual, or adverse conditions to reveal weaknesses and harmful behavior.
• reference-free evaluationReference-free evaluation assesses generated output from the input and output themselves without comparing them with reference texts.
• gan-based video generationGAN-based video generation is the creation of video content with a generative adversarial network.
• image-to-video synthesisImage-to-video synthesis is the generation of video content from a source image by modeling visual features and motion.
• video-to-video synthesisVideo-to-video synthesis is the generation of a new video from a source video while preserving aspects of its underlying structure.
• grounding layerA grounding layer is an application-architecture layer that acquires, prepares, represents, stores, and delivers external data for use by generative models.
• safetyAI safety concerns the prevention and mitigation of harmful, unreliable, or uncontrollable behavior by AI systems.
• vision modelA machine-learning model that processes or generates information represented in visual formats such as images.
• mode collapseMode collapse is a generative-model failure in which the model produces a narrow range of outputs repeatedly instead of representing the diversity of the data distribution.
• data augmentationData augmentation is the creation or transformation of training examples to increase dataset variety while preserving task-relevant information.
• deploymentA model deployment is a configured, addressable instance of a machine-learning model made available for serving requests.
• vector spaceA vector space is a mathematical set of vectors that supports operations such as addition and scalar multiplication.
• Chat Markup LanguageChatML is a markup format for representing chat messages and distinguishing their sources or roles for language-model processing.
• fill-in-the-blank attackA fill-in-the-blank attack prompts a language model to reconstruct forbidden content from partial words or omitted segments.
• many-shot jailbreakingMany-shot jailbreaking is a prompt attack that places numerous fabricated examples of an AI answering harmful questions before a final malicious request.
• obfuscationObfuscation is the transformation of content into a less recognizable representation while preserving information needed for later interpretation.
• output encodingOutput encoding is the conversion of generated data into a representation that prevents it from being interpreted as executable code in its destination context.
• principle of least privilegeThe principle of least privilege is a security principle that grants a system or component only the minimum access required to perform its intended functions.
• universal transferable suffixA universal transferable suffix is an adversarial prompt suffix designed to induce harmful behavior across multiple language models or prompts.
• harmHarm is an adverse effect or injury caused to individuals, groups, society, or other affected parties.
• user experienceUser experience is the overall perception and interaction quality a person has when using a product or system.
• testA test case is a defined combination of inputs, conditions, and expected results used to evaluate software behavior.
• bias detectionBias detection is the process of identifying systematic distortions or unfair tendencies in data, models, or outputs.
• Azure AI SearchAzure AI Search is Microsoft's cloud search service for indexing and retrieving content, including through vector and semantic search capabilities.
• Redis StackRedis Stack is a Redis distribution that bundles Redis with modules for additional data structures, search, and other capabilities.
• generative engineeringGenerative design is the computational creation of design alternatives from specified goals, constraints, and parameters.
• outpaintingOutpainting is an image-generation technique that extends an existing image beyond its original boundaries in a chosen direction.
• error analysisError analysis is the systematic examination of a model's mistakes to identify their types, causes, and recurring patterns.
• perplexityPerplexity is a language-model metric that quantifies uncertainty in predicting the next token of a sequence.
• prompt managementPrompt management is the systematic design, testing, deployment, refinement, and operational maintenance of prompts for language models.
• BERTScoreBERTScore is an automatic metric that evaluates generated text by comparing contextual semantic representations of candidate and reference text.
• alignment trainingAlignment is the process of shaping an AI model's behavior to conform to intended goals, values, or constraints.
• datasetA dataset is an organized collection of data records or examples assembled for analysis or machine-learning development.
• reward modelA reward model is a machine-learning model that assigns scores to outputs to estimate their quality or preference according to evaluation criteria.
• optimizationOptimization is the improvement of software or model behavior toward a desired objective such as efficiency, quality, or resource use.
• dynamic routingRouting logic is the decision mechanism that selects a model for an input or determines how outputs from multiple models are combined.
• ensembleA model ensemble is an architecture that combines multiple machine-learning models to improve overall capability, performance, or reliability.
• distillationKnowledge distillation is a model-compression technique in which a smaller student model learns to reproduce knowledge or behavior from a larger teacher model.
• few-shot chain of thoughtFew-shot chain of thought is a prompting technique that provides a small number of examples containing reasoning chains before asking the model to solve a new instance.
• encryptionEncryption is the transformation of readable data into a protected form that requires a key or equivalent mechanism to recover.
• authentication and authorizationAuthorization is the process of determining which resources or operations an authenticated entity may access.
• deduplicationDeduplication is the process of identifying and removing duplicate data records or content.
• hierarchical chunkingHierarchical chunking is a chunking method that divides text according to nested structural levels such as chapters, sections, and subsections.
• token-based splittingToken-based splitting is a chunking method that divides text according to a fixed number of tokens.
• BM25BM25 is a probabilistic ranking function that estimates document relevance to a query using term frequency, inverse document frequency, document length, and term-frequency saturation.
• lexical searchA sparse retriever is an information-retrieval system that matches queries and documents through sparse lexical representations and keyword-based statistics.
• sentence splitterA sentence-based splitter is a text-segmentation technique that divides text into chunks at sentence boundaries while preserving complete sentences.
• term frequency-inverse document frequencyTF-IDF is a statistical weighting scheme that scores a term according to its frequency in a document and its rarity across a document collection.
• ethical guidelineAn ethical guideline is a principle or rule that governs responsible development and use of an AI system.
• top-p filteringNucleus sampling is a text-generation technique that samples from the smallest dynamically selected set of tokens whose cumulative probability exceeds a specified threshold.
• Model ServingModel serving is the provision of a trained machine-learning model as an operational service that accepts inputs and returns predictions or generated outputs.
• transformer architectureThe Transformer architecture is a neural-network architecture based primarily on attention mechanisms for processing sequences.
• parameterized queryA prepared statement is a parameterized database query whose SQL structure is defined separately from the supplied data values.
• self-supervised learningA machine-learning technique that creates supervisory signals from the data itself to train a model without manually labeled examples.
• bias mitigationBias mitigation is the application of methods intended to reduce systematic unfairness or distortion in data, models, or outputs.
• OpenAI APIThe OpenAI API is a programming interface that provides access to OpenAI models and related services.
• machine learningA scientific discipline in which computers learn patterns from data to make predictions or decisions without being explicitly programmed for each task.
• pilot projectA small-scale implementation conducted to evaluate feasibility, learn from results, and guide broader deployment.
• exponential backoffA retry strategy that increases the delay between successive attempts, typically exponentially.
• finish reasonA finish reason is an API response field indicating why model generation ended or why the API returned the output it did.
• logit_biasLogit bias is a per-token generation setting that adds an offset to model logits before they are converted into probabilities.
• disaster recoveryDisaster recovery is the set of processes for restoring systems, data, and services after a catastrophic disruption.
• system integrationThe process of connecting software components or systems so they operate together and exchange capabilities or data.
• HarmBenchHarmBench is an open-source framework and benchmark for assessing the safety of large language models, especially their propensity to generate harmful content.
• robustnessRobustness is the ability of a machine-learning model to maintain acceptable behavior when inputs vary, are perturbed, or include difficult cases.
• TextAttackTextAttack is an open-source Python framework for generating adversarial examples, augmenting data, training NLP models, and evaluating their robustness.
• CodexCodex is a language model trained particularly on source code to generate and assist with programming code.
• byte pair encodingByte-pair encoding is a reversible, lossless compression and subword-tokenization algorithm that repeatedly merges frequent symbol sequences.
• tokenizerA tokenizer is a component that segments or encodes text into discrete units for processing by a language model.
• tiktokentiktoken is an OpenAI library for encoding text into the token sequences used by OpenAI language models.
• fallbackA fallback mechanism is an alternative execution or service path invoked when the preferred path fails or becomes unavailable.
• catastrophic forgettingCatastrophic forgetting is the loss of previously learned knowledge or capabilities when a model is trained on new data or tasks.
• distributional harmDistributional harm is harm whose effects are unequally distributed across a group or society rather than confined to one person.

These guides are generated from the text of Manning books by Manning's knowledge graph and link to the sections they draw on. The full books are available on liveBook and at manning.com.