Task 1.2: Perform data transformation, feature engineering, and pre-processing.

• ARobust Scaler

  Correct

  The Robust Scaler scales the column with statistics that are robust to outliers, so a handful of extreme values do not dominate how the ordinary values are spread.

• BMin Max Scaler

  Incorrect

  Min-max scaling maps the column to a fixed range from its minimum and maximum. A few huge values set the maximum, so nearly every other customer ends up packed near the bottom of the range.

• COrdinal encode

  Incorrect

  Ordinal encoding turns categories into integers. Monthly spend is already numeric, and the problem is its scale, not its type.

• DDrop missing

  Incorrect

  Drop missing removes rows that have missing values in the input column. It neither scales the feature nor addresses the outliers.

Distance-based models such as k-NN need features on comparable scales. When a feature has extreme values that must stay in the data, scale it with statistics such as the median and quantiles that are not pulled by those values.

• AOrdinal encode it, so each country becomes an integer from 0 to 39

  Incorrect

  Ordinal encoding suits categories with a natural order, such as education levels. For a linear model, integers would imply that country 39 is "larger" than country 1.

• BApply the Standard Scaler to the column

  Incorrect

  The Standard Scaler centers and scales numeric columns. A string category has to be encoded before any scaling makes sense.

• COne-hot encode it, one indicator column per country

  Correct

  Country is a nominal category with no inherent order. One-hot encoding gives every category its own indicator, so the model does not read any order or distance into the codes.

• DApply the Quantile Numeric Outliers transform to the column

  Incorrect

  This transform detects and fixes outliers in numeric features. The country column holds text categories, so there are no numeric outliers to fix and it still needs encoding.

Nominal categories (no order) are usually one-hot encoded; ordinal categories (an inherent order) can use ordered integer codes. The distinction matters most for linear and distance-based models.

• AHierarchical chunking with a parent and a child chunk size

  Correct

  Hierarchical chunking embeds small child chunks for precise matching and, at retrieval time, replaces them with their larger parent chunks so the model gets more complete context.

• BDefault chunking of about 300 tokens per chunk

  Incorrect

  Default chunking produces chunks of roughly 300 tokens that honor sentence boundaries. What is retrieved is what is returned, so there is no wider section around the match.

• CNo chunking, so each manual is one chunk

  Incorrect

  With no chunking each document becomes a single chunk. A whole manual as one embedding cannot match small passages precisely.

• DSemantic chunking with a higher breakpoint percentile threshold

  Incorrect

  Semantic chunking splits text where meaning changes, and a higher threshold gives fewer, larger chunks. It does not retrieve on small chunks and then return a larger parent.

Small chunks embed precisely; large chunks give context. Hierarchical chunking gets both by searching on child chunks and returning their parents. The documentation does not recommend it with an S3 vector bucket as the store, which is not the case here.

• AAn Amazon Macie sensitive data discovery job on the bucket

  Incorrect

  Macie discovers sensitive data in S3 objects and reports it as findings. It does not write a redacted copy of the text.

• BAn Amazon Comprehend sentiment analysis job on the tickets

  Incorrect

  Sentiment analysis returns whether text is positive, negative, neutral or mixed. It does not find or replace personal data.

• CA Comprehend PII redaction job that replaces each entity with its type

  Correct

  Comprehend redaction jobs read text from S3 and write a redacted copy. With MaskMode set to REPLACE_WITH_PII_ENTITY_TYPE, an entity such as "Jane Doe" becomes "[NAME]".

• DA Comprehend PII redaction job that masks each character with an asterisk

  Incorrect

  MASK mode replaces every character of an entity with the mask character, such as "**** ***". The text is protected, but the type label the team wants to keep is lost.

Comprehend can redact PII in batch from S3 in two ways: mask the characters or replace each entity with its type. Keeping the type preserves useful structure for training while removing the personal value.

• AA JSONL file of representative prompts, from which the teacher model generates the responses

  Correct

  Distillation input data is a collection of prompts. Bedrock sends them to the teacher model and uses the generated responses to fine-tune the student.

• BModel invocation logs from production traffic to the teacher model, stored in Amazon S3

  Correct

  If invocation logging is enabled, the job can use the logged prompts, or the logged prompt-response pairs when the teacher in the job matches the logged model.

• CA CSV file with one numeric feature per column and a label column

  Incorrect

  Distillation datasets are JSONL records in a conversation format for text-to-text models, not tabular feature files.

• DThousands of human-written answers, because distillation cannot use model-generated responses

  Incorrect

  The opposite is true: distillation is built on responses generated by the teacher model, so human-written answers are not required.

• EThe teacher model weights, exported from Amazon Bedrock to Amazon S3

  Incorrect

  Distillation needs prompts (and optionally logged responses), not the teacher's weights. The teacher is invoked by Bedrock during the job.

In distillation the teacher writes the training answers. You supply prompts, either directly as JSONL or from invocation logs, and Bedrock creates the synthetic data and fine-tunes the student on it.

• AProbabilistic encryption

  Incorrect

  Probabilistic encryption produces different ciphertext each time it is applied, so the same email would not match across the three datasets.

• BShuffling the values within the column

  Incorrect

  Shuffling moves values between rows, which breaks the link between each email and its own customer record, so joins return wrong matches.

• CDeterministic encryption

  Correct

  Deterministic encryption always produces the same ciphertext for the same value, so the encrypted column still joins across datasets, and holders of the key can decrypt it.

• DNulling out the column

  Incorrect

  Replacing the values with nulls hides them but leaves nothing to join on and nothing for an auditor to recover.

Pick the masking method from the requirements: joinable and reversible points to deterministic encryption; irreversible but joinable points to hashing; not needed at all points to nulling out or deleting the column.

• ANo chunking with the whole document as one chunk

  Incorrect

  No chunking makes each document one chunk, which loses the section-level granularity and cannot tag sections separately.

• BDefault chunking plus a metadata Lambda function

  Correct

  You can keep a built-in chunking strategy and provide a custom transformation Lambda function that receives the pre-chunked files and adds chunk-level metadata before the knowledge base continues processing.

• CSemantic chunking with a larger buffer size

  Incorrect

  Buffer size changes how semantic boundaries are found; it does not add section metadata to chunks.

• DA reranker model applied at query time

  Incorrect

  Reranking reorders results; it does not attach metadata for filtering.

Custom transformation Lambda functions can replace chunking entirely or just enrich built-in chunks with metadata.

• ARecreate the steps manually each month

  Incorrect

  Rebuilding the steps by hand is slow and error-prone when a saved recipe can be reused.

• BRun a profile job on each new monthly file

  Incorrect

  Profile jobs evaluate and summarize data; they do not apply transformation steps.

• CSave the steps as a recipe for a recipe job

  Correct

  DataBrew saves transformations as steps in a recipe, which you can update or reuse later with other datasets and deploy on a continuing basis.

• DUse a Glue crawler to transform the file

  Incorrect

  Crawlers catalog schemas; they do not transform data.

Make data preparation repeatable: capture steps once as a recipe and apply it to each new batch.

• AAn AWS Glue crawler on the topic

  Incorrect

  Crawlers catalog stored data; they do not consume or transform streams.

• BAn Athena CTAS query on the topic

  Incorrect

  Athena queries data at rest in S3; it does not read Kafka topics continuously.

• CA DataBrew profile job on the stream

  Incorrect

  Profile jobs summarize datasets; they do not run continuous stream transformations.

• DAn AWS Glue streaming ETL job

  Correct

  Glue streaming ETL jobs run continuously, consume sources such as Kinesis Data Streams, Apache Kafka and Amazon MSK, cleanse and transform the data, and load it into S3 data lakes.

For continuous Spark-based stream transformation without managing clusters, Glue streaming ETL is a managed option alongside Flink-based processing.

• ACATEGORICAL_MAPPING with an other value

  Correct

  CATEGORICAL_MAPPING maps categorical values to numeric or other values through a category map, and its "other" parameter replaces all non-mapped values.

• BONE_HOT_ENCODING of the education column

  Incorrect

  One-hot encoding creates a column per category and discards the order the team wants to keep.

• CBUCKETIZATION of the education column

  Incorrect

  Bucketization groups numeric values into ranges; the column holds text categories.

• DTOKENIZATION of the education column

  Incorrect

  Tokenization splits text into words; it does not assign ordered numbers.

Ordinal categories can be mapped to ordered numbers explicitly; nominal categories are usually one-hot encoded instead.

• AComprehend sentiment analysis on each message

  Incorrect

  Sentiment classifies opinion; it does not find personal data.

• BReal-time PII detection on each message, then masking

  Correct

  Comprehend can detect PII entities in a single document in real time, returning their types and offsets so the application can mask them as messages arrive.

• CComprehend topic modeling on the transcripts

  Incorrect

  Topic modeling groups documents by topics, does not redact PII, and is not available to new customers.

• DComprehend language detection on each message

  Incorrect

  Language detection identifies the language; it does not find or mask personal data.

• EAn asynchronous PII redaction job for the S3 batch

  Correct

  Comprehend asynchronous jobs read text from S3 and write a redacted copy, which fits the nightly batch.

Use synchronous PII detection for streaming, per-message masking, and asynchronous redaction jobs for large batches in S3.