Aura · Referência

Glossário de IA e LLM

Definições concisas do vocabulário usado na trilha.

Pesquisar no glossário

46 termos exibidos

AI agent: System that uses a model in a loop to observe, decide, call tools and update state toward a goal.
Algorithm: Finite procedure that transforms inputs into outputs through defined steps.
Artificial intelligence (AI): Field concerned with systems that perform tasks associated with perception, reasoning, learning, planning or action.
Attention: Mechanism that weights information from different positions when building a representation.
Benchmark: Standardized dataset or procedure used to compare systems, with limits set by its coverage.
Context window: Maximum working input and generated token span available to a model for one request.
Data: Recorded observations, examples or signals used to build, ground or evaluate a system.
Deep learning: Machine learning using multilayer neural networks to learn representations.
Direct preference optimization (DPO): Method that directly optimizes a language model from preferred and rejected response pairs.
Embedding: Vector representation in which learned relationships can be encoded geometrically.
Evaluation: Systematic measurement of behavior against defined tasks, cases, criteria and thresholds.
Fine-tuning: Additional training that adapts a pretrained model to data, behavior or tasks.
Foundation model: Broadly trained model intended to be adapted or used across many downstream tasks.
Generalization: Ability to perform on relevant inputs not seen during training.
Gradient: Direction and sensitivity of loss change with respect to model parameters.
Grounding: Connecting generation to supplied evidence, data, observations or verified tool results.
Guardrail: Technical or procedural control intended to prevent, detect or contain unwanted behavior.
Hallucination: Generated claim or detail that is unsupported by available evidence or is factually wrong.
Inference: Use of a trained model to produce an output for a new input.
Large language model (LLM): Large neural language model trained to predict or reconstruct tokens across broad data.
Latency: Elapsed time between a request and a useful response, often measured across percentiles.
Loss: Quantity training attempts to minimize to express prediction or objective error.
Machine learning (ML): Methods that fit behavior from data rather than specifying every rule manually.
Memory: Mechanism that persists or retrieves information beyond a model’s immediate context.
Model: Parameterized function or representation used to produce predictions, decisions or generated outputs.
Model Context Protocol (MCP): Open protocol for connecting AI applications to contextual resources, prompts and tools.
Multimodal: Able to process or generate more than one modality, such as text, image, audio or video.
Neural network: Layered parameterized computation inspired loosely by networks of simple connected units.
Overfitting: Fitting training examples too closely while performing poorly on relevant unseen cases.
Parameter: Value adjusted during training that shapes how a model transforms inputs.
Prompt: Instructions and context supplied to a generative model for a request.
Prompt injection: Untrusted content attempting to redirect a model or application away from intended instructions.
Quantization: Representing model values with lower precision to reduce memory or computation, with possible quality trade-offs.
Reinforcement learning (RL): Learning behavior through actions, observations and reward signals.
Retrieval-augmented generation (RAG): Retrieving external information and placing it in model context before generation.
RLHF: Reinforcement learning from human feedback, used to shape model behavior from preference signals.
Self-supervised learning: Learning signals derived from the data itself, such as predicting hidden or next parts.
Supervised fine-tuning (SFT): Fine-tuning on input and desired-output examples.
Supervised learning: Learning from examples paired with target labels or outputs.
Temperature: Sampling parameter that reshapes token probabilities; higher values generally increase variation.
Token: Unit processed by a language model; it may be a word, word part, punctuation or other symbol.
Tokenizer: Method that converts text or other input into token identifiers and back.
Tool calling: Model production of structured arguments for an application-controlled function or service.
Training: Process of adjusting model parameters to improve an objective on data or interaction.
Transformer: Neural architecture built around attention and feed-forward blocks, widely used in language models.
Vector database: System optimized to store vectors and retrieve items by similarity, often for semantic search.

Academia Aura

Aura · Referência

Glossário de IA e LLM

Definições concisas do vocabulário usado na trilha.

Pesquisar no glossário

46 termos exibidos

AI agent: System that uses a model in a loop to observe, decide, call tools and update state toward a goal.
Algorithm: Finite procedure that transforms inputs into outputs through defined steps.
Artificial intelligence (AI): Field concerned with systems that perform tasks associated with perception, reasoning, learning, planning or action.
Attention: Mechanism that weights information from different positions when building a representation.
Benchmark: Standardized dataset or procedure used to compare systems, with limits set by its coverage.
Context window: Maximum working input and generated token span available to a model for one request.
Data: Recorded observations, examples or signals used to build, ground or evaluate a system.
Deep learning: Machine learning using multilayer neural networks to learn representations.
Direct preference optimization (DPO): Method that directly optimizes a language model from preferred and rejected response pairs.
Embedding: Vector representation in which learned relationships can be encoded geometrically.
Evaluation: Systematic measurement of behavior against defined tasks, cases, criteria and thresholds.
Fine-tuning: Additional training that adapts a pretrained model to data, behavior or tasks.
Foundation model: Broadly trained model intended to be adapted or used across many downstream tasks.
Generalization: Ability to perform on relevant inputs not seen during training.
Gradient: Direction and sensitivity of loss change with respect to model parameters.
Grounding: Connecting generation to supplied evidence, data, observations or verified tool results.
Guardrail: Technical or procedural control intended to prevent, detect or contain unwanted behavior.
Hallucination: Generated claim or detail that is unsupported by available evidence or is factually wrong.
Inference: Use of a trained model to produce an output for a new input.
Large language model (LLM): Large neural language model trained to predict or reconstruct tokens across broad data.
Latency: Elapsed time between a request and a useful response, often measured across percentiles.
Loss: Quantity training attempts to minimize to express prediction or objective error.
Machine learning (ML): Methods that fit behavior from data rather than specifying every rule manually.
Memory: Mechanism that persists or retrieves information beyond a model’s immediate context.
Model: Parameterized function or representation used to produce predictions, decisions or generated outputs.
Model Context Protocol (MCP): Open protocol for connecting AI applications to contextual resources, prompts and tools.
Multimodal: Able to process or generate more than one modality, such as text, image, audio or video.
Neural network: Layered parameterized computation inspired loosely by networks of simple connected units.
Overfitting: Fitting training examples too closely while performing poorly on relevant unseen cases.
Parameter: Value adjusted during training that shapes how a model transforms inputs.
Prompt: Instructions and context supplied to a generative model for a request.
Prompt injection: Untrusted content attempting to redirect a model or application away from intended instructions.
Quantization: Representing model values with lower precision to reduce memory or computation, with possible quality trade-offs.
Reinforcement learning (RL): Learning behavior through actions, observations and reward signals.
Retrieval-augmented generation (RAG): Retrieving external information and placing it in model context before generation.
RLHF: Reinforcement learning from human feedback, used to shape model behavior from preference signals.
Self-supervised learning: Learning signals derived from the data itself, such as predicting hidden or next parts.
Supervised fine-tuning (SFT): Fine-tuning on input and desired-output examples.
Supervised learning: Learning from examples paired with target labels or outputs.
Temperature: Sampling parameter that reshapes token probabilities; higher values generally increase variation.
Token: Unit processed by a language model; it may be a word, word part, punctuation or other symbol.
Tokenizer: Method that converts text or other input into token identifiers and back.
Tool calling: Model production of structured arguments for an application-controlled function or service.
Training: Process of adjusting model parameters to improve an objective on data or interaction.
Transformer: Neural architecture built around attention and feed-forward blocks, widely used in language models.
Vector database: System optimized to store vectors and retrieve items by similarity, often for semantic search.