Page 134 - CyberSecurityEssentialsEnglish

TABLE 3-11: WHAT TO LOOK FOR IN PRODUCT DOCUMENTATION

Topic Critical reading questions

Which tasks and groups did the provider test? And which uses does
Intended use
it advise against?

Data and Which periods, languages, and domains are represented? And what
languages is not disclosed?

Does it show accuracy only, or also robustness, calibration, fairness,
Measures
efficiency, and cost?

Does the document concern the model, the system, or the
Version
application itself? And what is the date of the testing?

Are inputs used for training? What is the retention period? Where is
Privacy
the storage? And what options does an organization have?

Does it permit commercial use, modification, and redistribution?
License
And what are the terms for data and outputs?

Does the provider give notice when the model, the price, or the
Change
policy changes? And can a version be pinned?

Can the source be opened? And does it actually support the claim,
Citations
or is it only a general link?

Benchmarks and Contamination
A benchmark is a set of tasks, a measure, and a protocol. It may help with comparison, but

it does not equal performance in your work. You need to know the language, the domain,
the type of question, and the method of evaluation, and whether the result was zero-shot
or obtained with a special prompt or additional tools.

Benchmark contamination may occur when test questions or highly similar material leak into
the training or tuning data, so that a result appears higher than the true capacity to

generalize. A provider may also select the measures that highlight its strengths and omit
operating cost, stability, or privacy. Holistic evaluation addresses this shortfall through a
multi-scenario, multi-metric approach covering accuracy, calibration, robustness, fairness,

bias, toxicity, and efficiency [8].

? Question to consider: why did it win?

Did the product succeed on the test because its capability suits your task, or because the
test is close to its data, or because the comparison was set up in its favor? What evidence
could you reproduce yourself?

133