Task 1.2: Describe how various data types are used in gen AI and the business implications.

• AAudio recordings of customer service phone calls

  Correct

  Audio recordings have no predefined schema of rows and columns; like images, video and free text, they are unstructured data.

• BA table of account numbers, balances and opening dates

  Incorrect

  A table with fixed columns is structured data.

• CA spreadsheet of branch codes and opening hours

  Incorrect

  A spreadsheet with consistent columns is organized and easy to search, which makes it structured.

• DA ledger of card transactions with amounts and merchant codes

  Incorrect

  Transaction records with defined fields are classic structured data.

Structured data fits a predefined model of rows and columns; unstructured data such as audio, images, video and free text does not.

• ACompleteness

  Incorrect

  Completeness asks whether all the required information is present. Both systems have a value here; the values disagree.

• BFreshness

  Incorrect

  Freshness is about when data was last updated. The problem would be freshness if one figure were simply out of date, which the scenario does not say.

• CUniqueness

  Incorrect

  Uniqueness is a different dimension; it does not describe two systems reporting different values for the same fact.

• DConsistency

  Correct

  Consistency means the same value is the same across instances such as tables and systems; one month's customer count that differs between two systems is an inconsistency.

Google Cloud's data quality dimensions include completeness, consistency, freshness, validity, accuracy and uniqueness. Conflicting values across systems are a consistency issue, and a model trained on them inherits the conflict.

• AThe photos are converted into rows of a relational table first

  Incorrect

  Images do not have to become table rows to train an image model; the missing piece is the label, not a tabular format.

• BA sample of the photos is labeled as defective or not defective

  Correct

  A classifier that predicts a category learns from labeled examples, so a representative sample of the photos must first be annotated with the labels the model will predict.

• CThe photos are compressed so that more of them fit into one prompt

  Incorrect

  Prompt size is a concern for prompting a foundation model, not for training a custom image classifier.

• DThe model is given a larger context window to see more photos

  Incorrect

  Context windows apply to prompts sent to generative models; they do not replace the labeled examples a classifier needs.

Unlabeled data is raw information without tags; labeled data carries the answer the model should learn.

• AGive the assistant access to the current HR content

  Correct

  Data must be available and in the proper format to be useful. The answers are poor because the right data cannot be reached, so the first fix is access to the current source.

• BFine-tune the model on the old copy of the handbook

  Incorrect

  Tuning on outdated content would teach the model the wrong answers more firmly.

• CRaise the temperature so the answers sound more natural

  Incorrect

  Temperature changes how varied the wording is; it cannot supply facts the assistant has never seen.

• DSwitch to a newer model with a larger context window

  Incorrect

  A larger context window lets the model read more of what it is given; it still only gets the old copy.

Data accessibility is a precondition for quality: if the right data cannot be reached, no model choice or setting fixes the answers.

• AThe scans are older and must be refreshed before use

  Incorrect

  Freshness is about when data was last updated; both sources cover the same five years, so age does not set the contracts apart.

• BThe scans are unstructured, so data must be extracted first

  Correct

  Unstructured documents have no predefined schema; they need processing, such as OCR and field extraction with Document AI, to become structured, analyzable data.

• CThe scans include duplicates that must be removed first

  Incorrect

  Uniqueness problems come from repeated records; each contract was scanned once, so duplicates are not the extra work here.

• DThe scans are unlabeled and must be labeled by hand first

  Incorrect

  Labels matter for training supervised models; neither source is labeled, so labeling does not explain why the contracts take more work.

Structured data is ready to query; unstructured data has to be processed first, which is a cost to plan for. Freshness, uniqueness and labels are separate questions from structure.

• AConsistency

  Incorrect

  Consistency is about the same value matching across tables or systems; here the value is missing, not conflicting.

• BUniqueness

  Incorrect

  Nothing in the scenario points to a uniqueness problem; the values are missing.

• CCompleteness

  Correct

  Completeness asks whether the data contains all the information required for its intended purpose; a third of the rows lack a value the model needs.

• DValidity

  Incorrect

  Validity checks whether values follow the required format or range; empty cells are a completeness gap.

Missing values are a completeness problem, and a model trained on gaps can learn the wrong patterns for the affected regions.

• AFreshness of the transcripts it collects

  Incorrect

  Recent car-rental chats would still be about car rental; freshness does not fix the mismatch.

• BUniqueness of each individual transcript

  Incorrect

  Cleaning up the transcripts does not help when they come from the wrong business.

• CThe file format the transcripts are stored in

  Incorrect

  Format can be converted; content that does not match the task cannot.

• DRelevance of the data to the intended task

  Correct

  For generative AI, data quality includes relevance and alignment with the required outputs; transcripts from another business would teach the wrong content and tone.

More data is not better data. Tuning data should reflect the prompts and context the model will see in production.

• AReinforcement learning from human feedback on each example

  Incorrect

  Human feedback is used after pre-training, in an alignment phase where reviewers rate the model's responses; it does not label the pre-training data.

• BSelf-supervised learning, creating labels from the data itself

  Correct

  Foundation models are typically trained with self-supervised learning, which creates pseudo-labels from the data itself, for example by predicting masked or missing parts, so huge unlabeled datasets can be used.

• CSupervised learning on examples labeled by teams of people

  Incorrect

  The executive's premise is that nobody labeled the data; manual labels at that scale are expensive and slow, which is what self-supervised learning avoids.

• DGrounding with Google Search that adds web pages at runtime

  Incorrect

  Grounding connects a trained model to sources when it answers; the learning the executive asks about happens earlier, in pre-training.

Unlabeled data powers pre-training; labeled data and human ratings are most valuable later, for tuning, alignment and evaluating on a specific task.

• AAccuracy

  Incorrect

  Nothing suggests the numbers belong to the wrong customers; the problem is how they are written.

• BCompleteness

  Incorrect

  The values are present, so nothing is missing.

• CValidity

  Correct

  Validity checks whether data conforms to required standards such as format and acceptable ranges; a phone number in the wrong format is invalid data.

• DFreshness

  Incorrect

  Freshness measures when data was last updated, not how a value is written.

Format problems are validity problems. They matter because downstream tools and models expect one consistent format.