Page 116 - CyberSecurityEssentialsEnglish

automatically according to how sharp the distribution is: the set may be small when one
candidate dominates, and it widens when the probability is spread across several candidates.
These mechanisms may be used alone or together. The order in which temperature and the

filters are applied differs between systems and generation libraries, and it is not a fixed rule.
What matters in practice is that you record the settings you used and observe their effect

on your own cases, rather than assuming one universal order [5].

TABLE 3-3: CHOOSING GENERATION SETTINGS ACCORDING TO THE NATURE OF THE TASK

The case A reasonable starting setting What should be measured

Stability of fields, the
Field extraction or Low variety, with a structured
proportion of missing values,
disciplined rewriting output and validation rules.
and fidelity to the original.

Coverage of the points,
An executive Low to medium variety, with a
accuracy of figures, and
summary from a ban on additions beyond the
separation of text from
source source.
inference.

Medium or relatively high Diversity of alternatives,
Brainstorming or
variety, followed by a separate absence of repetition, and fit
creative alternatives
filtering stage. with the constraints.

Do not treat risk with Quality of evidence, failure
A high-impact
generation settings alone. Use cases, and abstention when
decision
sources, tools, and review. information is missing.

The Seed, Determinism, and Reproducibility
Some systems allow a random seed that fixes the starting point of the sampling process.
Fixing the seed and the settings may help produce closely similar results, but it does not

guarantee identical output if the model, the infrastructure, the search tool, or the order of
documents changes. Determinism means that the same conditions produce the same result;
it is stronger in calculations and rules than in composite generative systems.

Reproducibility in work and research requires a broader record. It covers the identity of the
system, meaning the model name and version, the service provider, and the date of the run.

It covers the inputs, meaning the user prompt, the system instructions where available, and
the files with their versions. It covers the settings, meaning the temperature, top-p, top-k,
and seed. And it covers the surrounding system, meaning the search and calculation tools,

the retrieval settings, the version of the knowledge base, the number of runs, and the region
or environment where that affects the service.

115