Skip to content
Cheat sheets

Generative AI Engineer Associate · Domain 5: Governance · 8% of the exam

Task 5.2: Use legal/licensing requirements for data sources to avoid legal risk; Recommend an alternative for problematic text mitigation in a data source feeding a GenAI application

Model and data licenses and terms, Marketplace listings and provenance, and removing or rewriting toxic, biased or sensitive source text before it reaches an index.

Study it

  • Model and data licenses

    Lesson coming

  • Mitigating problematic text in source data

    Partly covered by: Finding and Masking PII

Sample questions

Try each one before opening the answer. Every option is explained, with the Databricks documentation page that proves it.

Question 1 · choose 1

A team plans to distill a smaller in-house LLM by using thousands of rationales produced by Databricks-hosted LLM judges as training targets. The teacher model it also wants to query is listed on Applicable model terms with its own community license and acceptable use policy. Which plan avoids legal risk?

  1. AUse the judge rationales, since they are generated inside the Databricks security perimeter
  2. BRely on the Databricks agreement alone, since every hosted model is covered by the same terms
  3. CTrain on human-written labels, not judge outputs, and review the teacher model's listed terms first
  4. DUse any model on Applicable model terms freely, since Databricks has cleared the licenses
Show the answer and why

  • AUse the judge rationales, since they are generated inside the Databricks security perimeter

    Incorrect

    Where the content is processed does not change the restriction on using judge outputs to train an LLM.

  • BRely on the Databricks agreement alone, since every hosted model is covered by the same terms

    Incorrect

    Some models carry separate terms in addition to the Databricks agreement.

  • CTrain on human-written labels, not judge outputs, and review the teacher model's listed terms first

    Correct

    LLM judge outputs should not be used to train, improve or fine-tune an LLM, and customers are responsible for complying with each model's applicable terms.

  • DUse any model on Applicable model terms freely, since Databricks has cleared the licenses

    Incorrect

    The page lists terms for convenience; the customer remains responsible for compliance.

Databricks states that LLM judge outputs are for evaluating applications and should not be used to train or fine-tune an LLM. Models available through Model Serving may also be subject to their own licenses and acceptable use policies.

Databricks documentation

Task 5.2Report a problem with this question

Question 2 · choose 2

A community-forum corpus will feed a RAG assistant through a Delta Sync index. Some posts are pure abuse, and many helpful answers also contain insults. Policy says abusive language must never reach the index, but the helpful technical content must be kept. The cleanup runs in SQL in the ingestion pipeline. Which TWO steps fit? (Choose TWO.)

  1. ALabel each post with ai_classify and drop the ones labeled pure abuse before chunking
  2. BTell the model in the system prompt to ignore insults it finds in retrieved posts
  3. CRewrite the insulting helpful posts with ai_query so the insults are removed before chunking
  4. DStore a toxicity label with each chunk and exclude flagged chunks with an AI Search filter
  5. EExtract the insults into a separate column with ai_extract and leave the post text as is
Show the answer and why

  • ALabel each post with ai_classify and drop the ones labeled pure abuse before chunking

    Correct

    ai_classify assigns one of the labels you provide, and the label works as a filter that keeps abusive posts out of the chunk table.

  • BTell the model in the system prompt to ignore insults it finds in retrieved posts

    Incorrect

    The insults would still sit in the index and in the retrieved context; a prompt instruction does not remove them.

  • CRewrite the insulting helpful posts with ai_query so the insults are removed before chunking

    Correct

    ai_query applies a prompt to each row in batch, so the pipeline can rewrite a post to keep its technical answer without the insults.

  • DStore a toxicity label with each chunk and exclude flagged chunks with an AI Search filter

    Incorrect

    A query-time filter leaves the abusive text in the index, and it drops the helpful content along with the insults.

  • EExtract the insults into a separate column with ai_extract and leave the post text as is

    Incorrect

    Extraction adds structured fields; the original text, insults included, still gets chunked and indexed.

Filter unwanted documents out of the collection in the data pipeline, here by a label that ai_classify assigns, and transform the text worth keeping with ai_query, so that problematic content never reaches the source table of the index.

Databricks documentation

Task 5.2Report a problem with this question

Question 3 · choose 1

A team wants a "By request" dataset from Databricks Marketplace for a commercial RAG product. What does the access process involve?

  1. AInstant access, with no terms to accept
  2. BDescribe the intended use; the provider reviews it
  3. CCopying the data from a public download link
  4. DA support ticket that Databricks approves
Show the answer and why

  • AInstant access, with no terms to accept

    Incorrect

    Even instantly available products require accepting terms.

  • BDescribe the intended use; the provider reviews it

    Correct

    For By request listings you submit your name, company and intended use, accept the terms, and the provider reviews the request.

  • CCopying the data from a public download link

    Incorrect

    Marketplace access goes through the listing's process.

  • DA support ticket that Databricks approves

    Incorrect

    The provider, not Databricks support, reviews the request.

Marketplace products are free and instantly available (accept terms) or By request (submit intended use for provider review).

Databricks documentation

Task 5.2Report a problem with this question

Question 4 · choose 1

An admin disables Partner-powered AI features for legal reasons. The team still wants LLM judges in evaluation. What should it do?

  1. AStop using LLM judges entirely
  2. BRe-enable partner-powered features for each notebook that runs judges
  3. CUse judges only on synthetic data
  4. DConfigure the judges to use its own model
Show the answer and why

  • AStop using LLM judges entirely

    Incorrect

    Judges remain usable with a model you provide.

  • BRe-enable partner-powered features for each notebook that runs judges

    Incorrect

    The admin setting is not overridden per notebook.

  • CUse judges only on synthetic data

    Incorrect

    The data source does not change which model the judge calls.

  • DConfigure the judges to use its own model

    Correct

    Disabling partner-powered features stops judges from calling partner models, but you can still use judges with your own model.

LLM judges may use third-party services; admin settings for partner-powered AI and cross-Geo processing control where content is processed.

Databricks documentation

Task 5.2Report a problem with this question

Question 5 · choose 1

Data residency rules require evaluation content to stay in the workspace's Databricks Geo, so cross-Geo processing is disabled. What can happen when a built-in judge runs?

  1. AIt silently sends the content to a model in another Databricks Geo
  2. BIt skips scoring and reports a pass
  3. CIt returns a geo-restriction error if no in-Geo model exists
  4. DIt always uses a partner model abroad
Show the answer and why

  • AIt silently sends the content to a model in another Databricks Geo

    Incorrect

    Disabling cross-Geo processing prevents that.

  • BIt skips scoring and reports a pass

    Incorrect

    Judges do not report passes they did not compute.

  • CIt returns a geo-restriction error if no in-Geo model exists

    Correct

    With cross-Geo processing disabled, judges process content in the workspace Geo and return a geo-restriction error when no eligible in-Geo model is available.

  • DIt always uses a partner model abroad

    Incorrect

    Partner and Geo settings restrict, not force, external processing.

When cross-Geo processing is enabled, judges can process content in other Geos; when disabled, they are limited to the workspace Geo.

Databricks documentation

Task 5.2Report a problem with this question

Question 6 · choose 1

An old chain model logged with MLflow 1.15 lists the Anaconda defaults channel in conda.yaml. Legal is concerned about package licensing. What does Databricks suggest?

  1. ARe-log it with a conda_env that uses conda-forge
  2. BIgnore it, since the Databricks agreement covers Anaconda's channel terms
  3. CDelete conda.yaml from the artifacts
  4. DServe it on GPU to avoid conda
Show the answer and why

  • ARe-log it with a conda_env that uses conda-forge

    Correct

    Models logged before MLflow 1.18 may depend on the defaults channel; you can re-register with a new conda.yaml via conda_env.

  • BIgnore it, since the Databricks agreement covers Anaconda's channel terms

    Incorrect

    Use of Anaconda channels is governed by Anaconda's terms.

  • CDelete conda.yaml from the artifacts

    Incorrect

    The model environment needs a valid dependency file.

  • DServe it on GPU to avoid conda

    Incorrect

    Compute type does not change the environment's channels.

Anaconda changed its terms for anaconda.org channels. MLflow 1.18 and later default to conda-forge; older models can be re-logged.

Databricks documentation

Task 5.2Report a problem with this question

Practice domain 5 →Practice all domains →