Page 149 - CyberSecurityEssentialsEnglish

evaluation set, repeated runs, an error taxonomy, a total cost, and a test designed
specifically for Arabic.

A good decision remains provisional. Models, interfaces, prices, policies, and knowledge
bases all change. The selection process therefore does not end at purchase or adoption; one

cycle ends and a cycle of monitoring and re-evaluation begins. In the next chapter the book
moves from the question “which tool is right?” to the question “how do I use it safely without

losing control?”

Common
Term Plain-language explanation
abbreviation

A learned or rule-based component that converts an input
Model -
into a prediction, content, or representation.

A model trained on broad data that can be adapted to many
Foundation model -
tasks.

A model trained on very large amounts of text that
Large language
LLM produces language by predicting what comes next; not a
model
synonym for all of artificial intelligence.

Small language A smaller model that may be faster, cheaper, and better
SLM
model suited to local operation.

A unit of text the model processes; it may be a word or part
Token -
of one.

Optical character Turning the image of a document into text a computer can
OCR
recognition process.

Selecting the most probable token at every step, which
Greedy decoding -
raises consistency and lowers variety.

A setting that reshapes the probabilities of selection and
Temperature -
usually affects variety.

Nucleus sampling: selecting from the smallest set whose
Top-p -
cumulative probability reaches a defined limit.

Restricting the choice to a fixed number of the most
Top-k -
probable candidates.

A value that fixes the starting point of sampling and helps
Seed -
make a run more repeatable.

The space the system can process in the current interaction,
Context window -
including input and output.

Selecting, ordering, and compressing what enters the
Context engineering -
context window, and deciding what to retrieve on demand.

148