Task 1.3: Validate data quality and manage bias.

• AAn AWS Glue crawler that runs on the output prefix after each job

  Incorrect

  A crawler infers schemas and adds or updates tables in the Data Catalog. It does not test the values in the data against rules.

• BAWS Glue Data Quality rules in DQDL, evaluated in the ETL job

  Correct

  Glue Data Quality evaluates rules such as IsComplete, IsUnique and range checks written in the Data Quality Definition Language, both on Data Catalog tables and inside Glue ETL jobs, and the job can act on the results.

• CAmazon CloudWatch alarms on the Glue job metrics

  Incorrect

  Glue job metrics describe the job's execution, such as memory and data moved, not whether each value in the table is valid.

• DAn Amazon Macie job on the output bucket

  Incorrect

  Macie discovers sensitive data such as PII in S3 objects. It does not check completeness, uniqueness or value ranges.

Glue Data Quality, built on the open-source Deequ framework, brings declarative data checks into the pipeline so bad data can be stopped or quarantined before it is used for training.

• AKeep accuracy as the objective metric for model selection

  Incorrect

  On a dataset this imbalanced, predicting "not fraud" every time already scores 99%. A metric such as balanced accuracy, F1 or recall shows whether fraud is caught.

• BUndersample the test set too, so it has the same class ratio as training

  Incorrect

  Changing the test set's class mix makes the evaluation unlike production traffic, so the reported numbers stop predicting real performance.

• COversample the fraud class in the training split, for example with SMOTE

  Correct

  Balancing operations such as random oversampling or SMOTE add minority-class samples, which helps a binary classifier learn the rare class. Only the training data should be balanced, so evaluation still reflects real traffic.

• DSet the scale_pos_weight hyperparameter to weight the positive class more

  Correct

  scale_pos_weight controls the balance of positive and negative weights in XGBoost and is meant for unbalanced classes.

• EShuffle the rows before splitting the data

  Incorrect

  Shuffling changes which rows land in each split, but every split still has about 1% fraud, so the imbalance is untouched.

Fix imbalance where the model learns (resampling the training data or reweighting the classes) and judge the result with metrics that do not reward always predicting the majority class.

• AA stratified split on the demand column

  Incorrect

  A stratified split keeps the proportion of each target value the same in every split. Rows are still drawn from the whole time range, so future days still inform training.

• BA randomized split with a different random seed

  Incorrect

  A different seed gives a different random sample, but every random split still mixes future days into the training set.

• CA split by key on store_id, so each store is in one split

  Incorrect

  Split by key keeps each store in only one split. Every store's history still spans all dates, so the time leakage remains.

• DAn ordered split that holds out the last 20% of days

  Correct

  An ordered split keeps the sequence of the observations: the first 80% train and the last 20% test. Validating on later days than training mirrors how the forecast is used.

For time-ordered data a random split leaks the future into training and inflates validation scores. Hold out the most recent period instead.

• AA dropDuplicates() step on all columns in a Spark job

  Incorrect

  Dropping duplicate rows removes rows whose values are identical. Records that differ by a spelling or an abbreviation are kept as two customers.

• BA DQDL IsUnique rule on the name column

  Incorrect

  IsUnique checks whether values in a column are unique. It reports a problem but neither decides which differently spelled records refer to the same person nor merges them.

• CAn AWS Glue crawler on both sources

  Incorrect

  A crawler catalogs schemas. It does not compare records with each other.

• DA FindMatches transform in an AWS Glue job

  Correct

  FindMatches is a machine learning transform that finds duplicate or matching records even when they have no common unique identifier and no fields match exactly. You teach it with labeled examples and call it from a Spark-based Glue job.

Exact deduplication only catches identical rows. Fuzzy record matching across sources without a common key is what FindMatches is for.

• AImpute the mean of the column, then drop the original column

  Incorrect

  The mean of a long-tailed column is pulled toward its extreme values, and dropping the original column throws the feature away.

• BImpute the median and add a missing-value indicator column

  Correct

  Median imputation fills gaps with a value that extreme values barely move, and the "add indicator for missing" transform creates a Boolean column so the model can still use the fact that a value was missing.

• CDrop every row that has a missing debt_ratio value

  Incorrect

  Drop missing removes 12% of the rows, which the team wants to avoid, and loses the signal in the missingness itself.

• DFill every missing value with zero

  Incorrect

  Zero is a real, valid-looking debt ratio, so the model cannot tell filled rows from customers who truly have no debt.

Imputation keeps rows; the median resists extreme values; an indicator column keeps the information that the value was missing. All three are standard cleaning steps in Canvas data preparation.

• AScreen every pair with ApplyGuardrail and a configured guardrail

  Correct

  ApplyGuardrail assesses any text against a configured guardrail's content filters, denied topics, PII detectors and word lists without invoking a foundation model, so it can screen a dataset before training.

• BRun AWS Glue Data Quality IsComplete and IsUnique rules on both columns

  Incorrect

  These rules check that values are present and unique. They say nothing about whether the text is insulting or touches a denied topic.

• CRun an Amazon Macie sensitive data discovery job on the dataset bucket

  Incorrect

  Macie looks for sensitive data such as PII in S3 objects. It does not detect hate speech or company-specific denied topics.

• DRun Amazon Comprehend sentiment analysis and drop every negative pair

  Incorrect

  Sentiment shows whether text is positive or negative. Polite text can still cover a denied topic, and negative text is not necessarily harmful.

Training data integrity for foundation models includes content safety. Because ApplyGuardrail is decoupled from model invocation, the same guardrail used at runtime can screen fine-tuning data in advance.

• AA DataBrew recipe job with no steps

  Incorrect

  A recipe job applies transformations; without steps it produces a copy, not a profile.

• BA DataBrew profile job on the dataset

  Correct

  Profile jobs run a series of evaluations on a dataset and output the results to S3, helping you understand the data before preparing it; profile jobs can also be configured to detect PII.

• CAn Amazon Athena CTAS query on the data

  Incorrect

  CTAS converts query results into new tables; it does not profile column distributions or detect PII.

• DAn S3 Inventory report on the bucket

  Incorrect

  An inventory lists objects and their properties, not the contents of the data.

Profile before you prepare: statistics, missing values and sensitive fields guide which cleaning steps are needed.

• AFail the whole job on any rule failure

  Incorrect

  Stopping the job also blocks all the good rows, which the team wants to keep loading.

• BAn AWS Glue crawler run after each job

  Incorrect

  A crawler updates schemas; it does not identify failing records.

• CGlue DQ record-level results to quarantine rows

  Correct

  AWS Glue Data Quality helps identify the exact records that caused quality scores to go down, so they can be quarantined and fixed while good data continues through the pipeline.

• DAn S3 Lifecycle rule on the output prefix

  Incorrect

  Lifecycle rules move or expire objects by age; they do not check data quality.

Quarantining bad records keeps pipelines flowing while problems are investigated, instead of choosing between bad data and no data.

• AImpute missing for the categorical column

  Incorrect

  For categorical data, Impute missing uses the most frequent value, which guesses a region instead of marking the value as unknown.

• BDrop missing on the region column

  Incorrect

  Drop missing removes the rows, which loses data instead of labeling it.

• CRobust Scaler on the region column

  Incorrect

  Scalers transform numeric data; region is a text category.

• DFill missing with the value "Unknown"

  Correct

  Impute missing fills categorical gaps with the most frequent value; to use a custom string instead, the documentation points to the Fill missing transform.

"Missing" can itself be informative; filling with an explicit placeholder category keeps that information.

• AConcatenate, removing duplicates afterward

  Correct

  Concatenate appends the rows of one dataset to another, and you can select an option to remove duplicates after concatenation.

• BAn inner join of the files on customer_id

  Incorrect

  A join combines columns from two tables side by side; it does not stack the monthly rows into one dataset.

• CA randomized split of the two files

  Incorrect

  Splitting divides data into subsets; it does not combine files.

• DSimilarity encode on the customer_id column

  Incorrect

  Similarity encoding creates embeddings for categories; it does not merge or deduplicate datasets.

Stack datasets with concatenation, combine columns with joins, and remove exact duplicates as part of the merge when sources overlap.

• AIncrease the number of training epochs

  Incorrect

  The job fails while parsing the file, before training settings matter.

• BOne JSON object per line, not an array

  Correct

  The error indicates the file is not valid JSONL. Customization datasets are .jsonl files in which each line is a JSON object for one record.

• CCompress the training file with gzip

  Incorrect

  Compression does not turn a JSON array into JSON Lines.

• DGrant the role kms:Encrypt on the bucket key

  Incorrect

  A parsing error points to the file format, not to key permissions.

Validate training files before submitting jobs: correct JSONL structure first, then field names and size limits.