Task 1.3: Data Preparation for ML and AI - Validate data quality and manage bias

Skill 1.3.6: Validate AI training data integrity (for example, prompt-response pair validation, content safety screening)

1. Why AI Training Data Integrity Matters

For generative AI and supervised fine-tuning, the training dataset is often built from prompt-response pairs. These pairs teach the model how to respond to user inputs. If the data is malformed, inconsistent, private, offensive, or misaligned, the model can learn these defects and reproduce them at inference time.

Key principle:

"Content that goes in can come back out in a model response."

Therefore, training data integrity is not only about schema completeness. It also includes:

Checking that each prompt has a valid, corresponding response
Ensuring the response genuinely answers the prompt
Detecting duplicate or near-duplicate records
Screening for toxic, sensitive, or denied content
Removing personally identifiable information (PII)
Confirming formatting and role structure are consistent

2. Prompt-Response Pair Validation

Prompt-response pair validation is both structural and semantic.

A well-formed pair should have:

Component	Validation Question
Prompt presence	Is the prompt field present and non-empty?
Response presence	Is the response field present and non-empty?
Role structure	Are roles such as `user`, `assistant`, `system` correctly assigned?
Formatting	Are special tokens, separators, and turn boundaries consistent?
Alignment	Does the response actually answer the prompt?
Completeness	Are there any truncated responses or missing fields?
Uniqueness	Are there duplicate or near-duplicate pairs that could over-represent one pattern?
Safety	Does the prompt or response contain unsafe or policy-violating content?

2.1 Structural Validation

Structural validation checks whether every record is complete and formatted correctly. This is similar to validating any dataset, but applied to generative AI data.

Common structural defects:

Defect	Example	Risk
Missing response	`{"prompt": "What is the return policy?", "response": ""}`	Model learns not to answer
Missing prompt	`{"response": "Our return policy is 30 days."}`	Model learns meaningless output
Truncated response	`"response": "The warranty covers"`	Model learns incomplete text
Wrong field order	Prompt and response content swapped	Model learns reversed behavior
Repeated pairs	500 identical question-answer copies	Overfitting and bias toward one pattern
Inconsistent role names	`"human"` / `"bot"` instead of expected format	Confuses model during fine-tuning

AWS tools for structural checks:

AWS Glue Data Quality can evaluate completeness, uniqueness, row count, and schema rules.
AWS Glue DataBrew can profile the dataset and surface missing values, duplicates, or inconsistent formats.
Amazon SageMaker Data Wrangler can be used interactively to clean and transform the pairs.

Example rule concept without code:

Completeness of prompt must be greater than 99%
Completeness of response must be greater than 99%
Uniqueness of combined prompt + response must be greater than 95%
response length must be between 1 and 2000 characters

2.2 Semantic Alignment Validation

A prompt and response can both be present and well-formatted, yet still misaligned.

Example:

Prompt: "How do I reset my password?"
Response: "Shipping takes 5–7 business days."

This pair is structurally valid but semantically invalid. The model would learn the wrong association.

Approaches:

Approach	Description
Human review	Use human reviewers to judge whether the response answers the prompt
LLM-as-judge	Use a stronger model to score relevance and correctness of each pair
Embedding similarity	Compare prompt and response embeddings to detect suspiciously unrelated pairs
Rule-based checks	Check for expected keywords, answer length, or response category

When to use human review:

Responses require domain expertise
There is ambiguity or disagreement in automated results
The cost of an incorrect response is high, such as in healthcare or finance

AWS service:

Amazon SageMaker Ground Truth can create human review workflows.
Amazon Augmented AI (A2I) is more commonly used for inference-time human review, while Ground Truth is used for labeling and training data.

3. Content Safety Screening

Content safety screening prevents unsafe, offensive, private, or policy-violating content from entering the training dataset.

A generative model trained on toxic or sensitive content can:

Reproduce toxic language
Leak personally identifiable information
Generate illegal or harmful instructions
Produce brand-damaging responses
Violate regulatory requirements

3.1 Safety Categories to Screen

Safety Category	Examples
Toxicity / hate speech	Insults, slurs, harassment
Violence	Descriptions of harm, graphic content
Sexual content	Explicit adult content
Illegal activity	Instructions for illegal acts
PII	Emails, phone numbers, addresses, SSNs
Sensitive financial/health data	Account numbers, medical records
Denied topics	Topics banned by company policy
Offensive words	Profanity, slurs
Prompt injection / adversarial content	Instructions attempting to override model behavior

3.2 AWS Tools for Content Safety

AWS Service / Feature	Purpose
Amazon Bedrock Guardrails	Block or filter content using configurable policies
Amazon Comprehend	Detect PII entities in text
AWS Glue Data Quality	Validate schema, completeness, uniqueness, and allowed values
Amazon SageMaker Ground Truth	Human review for nuanced safety judgments
Amazon CloudWatch	Monitor validation job failures and content screening metrics

Amazon Bedrock Guardrails filters relevant to training data preparation:

Filter Type	What It Does
Content filters	Block harmful content across categories such as hate, violence, sexual content, and insults
Denied topics	Block topics that are not allowed by company policy
Word filters	Block specific words or profanity
Sensitive information filters	Block or mask PII and sensitive data
Contextual grounding filters	Validate that responses are grounded in the correct source material and relevant to the query

Important:

Bedrock Guardrails can be used to screen both prompts and responses.
The filters are configurable; thresholds and categories can be adjusted.
Screening should happen before the data is used for customization or fine-tuning.

4. End-to-End Validation Flow

flowchart TD
    A[Raw prompt-response dataset] --> B[Structural validation]
    B -->|Pass| C[Semantic alignment validation]
    B -->|Fail| Q[Quarantine or repair]
    C -->|Pass| D[Content safety screening]
    C -->|Fail| Q
    D -->|Pass| E[Approved AI training dataset]
    D -->|Fail| Q
    Q --> F[Review, clean, redact, or discard]

5. Common AWS Solutions for Training Data Integrity

Integrity Task	Recommended AWS Approach
Detect missing prompt/response fields	AWS Glue Data Quality completeness rules
Detect duplicate prompt-response pairs	AWS Glue Data Quality uniqueness rules
Profile data before cleaning	AWS Glue DataBrew profiling job
Clean malformed records	SageMaker Data Wrangler or AWS Glue DataBrew
Detect PII in text	Amazon Comprehend
Block toxic or denied content	Amazon Bedrock Guardrails
Review ambiguous semantic alignment	SageMaker Ground Truth human review
Monitor validation pipeline health	Amazon CloudWatch

6. Scenario-Based Examples

Scenario 1: Customer Service Fine-Tuning Dataset

A company is preparing a dataset of customer support interactions to fine-tune a generative AI assistant. The data contains the following:

Several records with empty response fields
Many duplicated questions with identical answers
Some responses containing customer email addresses

Question:
Which combination of AWS capabilities should the team use to improve data integrity before fine-tuning?

Options:

A. Use AWS Glue Data Quality for completeness and uniqueness rules, and use Amazon Comprehend to detect PII
B. Shuffle the dataset and increase the number of training epochs
C. Use Amazon SageMaker Feature Store to store prompt-response pairs
D. Convert the dataset to Apache Parquet format

Correct Answer: A

Explanation:
Structural problems like missing responses are detected with completeness rules. Duplicate prompt-response pairs are detected with uniqueness rules. Customer email addresses are PII and should be detected using Amazon Comprehend or blocked using Bedrock Guardrails sensitive information filters.

Why the other options are incorrect:

Shuffling does not repair missing fields, remove duplicates, or remove PII.
Feature Store is for feature storage and reuse, not for validating content integrity.
Columnar format improves storage and query performance, not data quality.

Scenario 2: Safe Language for a Children’s Product

A company is building a generative AI feature for a children’s educational application. The fine-tuning dataset contains some responses with profanity and violent language.

Question:
Which AWS service feature is most appropriate for screening this dataset before training?

Options:

A. AWS Glue Data Quality rules
B. Amazon Bedrock Guardrails word filters and content filters
C. Amazon SageMaker Data Wrangler
D. Shuffling the records

Correct Answer: B

Explanation:
Bedrock Guardrails provides word filters to block profanity and content filters to block harmful content such as violence. These filters can be applied to training data before customization to reduce the likelihood of the model learning unsafe language.

Why the other options are incorrect:

Glue Data Quality checks schema, completeness, and uniqueness, not toxic content.
Data Wrangler cleans and transforms data but does not provide content safety policies.
Shuffling only reorders rows and does not remove harmful language.

Scenario 3: Responses That Do Not Answer the Prompt

A healthcare team is fine-tuning a medical question-answering model. After structural validation, they find that many records have both prompt and response fields filled, but some responses do not answer the prompt.

Example defect:

Prompt: "What are the early symptoms of diabetes?"
Response: "Schedule an appointment with your provider for lab work."

Although the response is safe and non-empty, it does not answer the question.

Question:
What type of validation is needed to catch this issue?

Options:

A. Semantic alignment validation
B. Schema completeness validation
C. Duplicate detection
D. Column profiling

Correct Answer: A

Explanation:
Semantic alignment validation checks whether the response actually answers the prompt. This can be done through human review, an LLM-as-judge, or embedding similarity.

Why the other options are incorrect:

Completeness validation only checks if fields are populated.
Duplicate detection only checks repeated records.
Column profiling summarizes distributions and missing values but does not evaluate prompt-response meaning.

7. Exam Tips and Traps

✅ Tips

Think of integrity checks in layers
Structure → Semantic alignment → Safety screening
A complete dataset can still be semantically misaligned or unsafe.
Use AWS Glue Data Quality for declarative rules
Completeness, uniqueness, row count, schema, and allowed value ranges.
Rules are repeatable and evaluated on every pipeline run.
Use Amazon Comprehend for PII
It detects PII entities in text, such as names, emails, phone numbers, and addresses.
It supports both English and Spanish text.
Use Bedrock Guardrails for content safety
Word filters block profanity.
Content filters handle toxicity, violence, sexual content, and more.
Sensitive information filters can block or mask PII.
Denied topics enforce custom business policies.
Remember that screening happens before training
Unsafe content in the training set can later appear in model responses.
Combine automated checks with human review
Automated tools catch obvious structural and safety problems.
Human review is better for semantic alignment and nuanced policy decisions.
Monitor failures with CloudWatch
A silent validation failure can allow bad data into training.

❌ Traps

Do not confuse shuffling with integrity repair
Shuffling changes row order.
It does not fix missing fields, malformed pairs, PII, toxicity, or misalignment.
Do not use class imbalance techniques for integrity problems
Oversampling, undersampling, and class weighting fix label imbalance.
They do not fix malformed prompt-response pairs or unsafe content.
Do not assume completeness rules detect unsafe content
A field can be non-empty and still contain toxic or private data.
Safety screening is a separate step.
Do not choose Amazon Macie for prompt-response PII screening
Amazon Macie discovers PII in S3 objects for data security and compliance.
Amazon Comprehend is better for detecting PII entities directly in text records.
Do not treat structural validity as semantic alignment
Just because a response exists does not mean it answers the prompt.
Do not ignore near-duplicates
Near-duplicate prompt-response pairs can cause the model to overfit to a narrow pattern.
Do not apply Bedrock Guardrails only at inference
Exam questions may present an inference-only safety solution. The correct answer for data preparation should include screening the training dataset.
Do not rely on overall accuracy for generative AI validation
Accuracy is not the primary measure for prompt-response data quality.
Integrity is measured by completeness, alignment, safety, and consistency.

8. Quick Summary

Question	Correct Mental Model
Is the record complete?	Use AWS Glue Data Quality completeness rules
Is the record duplicated?	Use uniqueness rules or deduplication
Is PII present?	Use Amazon Comprehend or Bedrock Guardrails sensitive info filters
Is the content toxic or denied?	Use Amazon Bedrock Guardrails content, word, and denied topic filters
Does the response answer the prompt?	Use human review, LLM-as-judge, or embedding similarity
Should I fix underrepresentation with shuffling?	No, shuffling only reorders existing records
When should safety screening happen?	Before training, not only after deployment

9. Final Word

AI training data integrity for generative AI is a multi-step validation process:

Confirm that prompt-response pairs are structurally complete and unique.
Confirm that the response semantically answers the prompt.
Screen the content for privacy, toxicity, and policy violations.
Monitor the validation pipeline to catch failures.

The correct AWS solution usually combines AWS Glue Data Quality, Amazon Comprehend, Amazon Bedrock Guardrails, and SageMaker Ground Truth, depending on the specific defect being addressed.