NCA-GENL Flashcards

Generative models create new outputs from learned distributions, while discriminative models classify or score existing inputs.

Supervised uses labels, unsupervised finds structure, and self-supervised builds labels from raw data (the core of LLM pretraining).

The model learns useful internal features (embeddings/hidden states) that can transfer across many tasks.

Broad pretrained models that can be adapted with prompting, RAG, or fine-tuning for specific use cases.

Performance generally improves with more parameters, data, and compute, but with cost and diminishing-return tradeoffs.

Reusing pretrained model knowledge to reduce training cost/time for downstream tasks.

Stacked linear and nonlinear layers transform inputs into progressively richer features.

Nonlinearities that let networks model complex patterns; GELU/ReLU are common in modern transformer stacks.

Cross-entropy is typical for token prediction/classification; MSE is common for regression targets.

Algorithm that computes gradients of loss with respect to model parameters.

Optimi

Lets each token weigh relevance of other tokens in the sequence.

Uses multiple attention projections so the model can learn different relation types in parallel.

Adds token-order information because attention alone is permutation-invariant.

Encoder-only is strong for understanding, decoder-only for generation, and encoder-decoder for input-to-output transformation.

Per-token nonlinear layers that refine representations after attention.

Skip paths that preserve signal and stabili

Large-scale self-supervised training to build general language capability.

Additional training to adapt the base model to a target task/domain.

Fine-tuning with labeled prompt-response examples.

Trains the model to better follow user instructions and response formats.

Reuse pretrained knowledge instead of training from scratch.

Selecting, cleaning, balancing, and deduplicating training data.

Normali

Combines retrieval from external knowledge with model generation.

Dense numeric representations capturing semantic meaning.

Models that convert text (or multimodal data) into vectors.

Stores/indexes embeddings for nearest-neighbor search.

Retrieves by meaning similarity rather than exact keyword matches.

Distance/similarity formulas used for vector ranking.

Methods for splitting documents into retrievable units.

Adds context continuity between adjacent chunks.

Restricts retrieval by fields like source, date, tenant, or policy.

Combines lexical and vector retrieval to improve recall/precision.

Secondary ranking stage that improves relevance among initial retrieval results.

Producing answers that are explicitly supported by retrieved evidence.

Building and maintaining curated source corpora for retrieval.

End-to-end ingest, parse, embed, and index workflows for searchable knowledge.

Updating most or all model parameters for maximum task adaptation.

Parameter-efficient fine-tuning; updates only a small subset of weights.

Low-rank adapters inserted into target layers for efficient adaptation.

Small trainable modules added between fro

Reinforcement learning from human feedback to align outputs with user preferences.

Trains a model to score response quality from preference data.

Review this concept in the chapter content.

Measures how well a model predicts token sequences (lower is generally better).

N-gram overlap metric often used for translation quality.

Overlap metric family commonly used for summari

Systematic skew in outputs caused by data, training, or deployment patterns.

Ensuring comparable quality/treatment across different groups and contexts.

Identifying harmful or abusive language in input/output streams.

Blocking/redacting policy-violating or unsafe content.

Rule and policy layers constraining model behavior and tool actions.

Untrusted content attempting to override instructions and controls.

Attempts to bypass safety policies through adversarial prompt patterns.

Protection of sensitive information in training, retrieval, and inference.

Ownership, approvals, auditability, and change-control for AI systems.

Legal and regulatory requirements tied to industry/jurisdiction.

Transparency, accountability, fairness, privacy, and harm reduction in design and operations.

Models that process and generate across text, image, audio, or video.

Architectures that jointly reason over visual and textual inputs.

Systems that generate images from natural-language prompts.

Generative models that iteratively denoise latent noise into outputs.

Generator-discriminator training framework for synthetic data creation.

Shared vector spaces connecting semantics across modalities.

Generating descriptive text from image content.

Producing temporally coherent sequences from prompts/conditions.

Review this concept in the chapter content.

Automated data movement/processing from source to model-ready artifacts.

Extract, transform, load steps for structured data preparation.

Creating useful input signals from raw data.

Generating, storing, and refreshing vector representations.

Creating supervised targets and annotation quality controls.

Immutable dataset snapshots for reproducibility and audit.

Policies for data quality, access, lineage, and stewardship.

Tracking model artifacts across training and deployment cycles.

Logging parameters, metrics, artifacts, and outcomes per run.

Operational practices for reliable ML/LLM delivery.

Automated validation, packaging, and release processes for model systems.

Continuous quality, latency, cost, and safety monitoring.

Detecting shifts in input data or model behavior over time.

Using user/system signals to drive iterative improvements.