Analysis package (clean)
A service analysis dataset: record_id, period, group, value, unit, status, sensitive_flag, source_version. Used in the Guided Lab.
DS 202 Practice Center
Use the synthetic Analysis Package dataset to practice supplying schema context, hunting AI-hallucinated columns, reviewing silent errors, verifying in layers, logging provenance, and deciding whether sensitive data may be shared with AI.
Info
The DS 202 practice data is synthetic teaching data; every service group, record, and value is illustrative and does not represent any real organization or person.
Warning
Use only the synthetic files provided by the course. Unless authorized, never paste confidential institutional or company data, personal data, or other sensitive information into an AI tool; check your institution's policy before working with real operational data.
Download the practice files
Analysis package (practice, with challenges)
The same schema, with deliberate judgment challenges: a column that looks useful but is actually hallucinated (not part of the canonical schema), analysis flaws from not filtering by status, a duplicated record_id, and at least one row where sensitive_flag is true. Used in the Practice Center and the Independent Lab.
Data dictionary (Traditional Chinese)
Column types, units, grain, allowed values, and provenance for the analysis package file (Traditional Chinese).
Data dictionary (English)
English data dictionary; columns and units match the Chinese version.
Guided Lab starter notebook
A starter notebook with instructions and items to complete, runnable top-to-bottom from a fresh kernel, practicing context-setting, verifiable prompts, and read-before-run.
Independent Lab starter notebook
A starter notebook and checklist template for the AI-Assisted Workflow + Verification Record.
Practice tasks
Practice A: Schema Context Prompt
For the practice data, write a complete data context and use it to build a structured analysis prompt so AI does not have to guess the columns itself.
Files used: Analysis package (practice, with challenges), Data dictionary (Traditional Chinese)
Steps
- List every column's type, allowed values, and grain, marking forbidden assumptions (for example, never assume every row's status is already validated).
- Write down every value the unit column can take, and what must happen before aggregating across units.
- Write an analysis prompt using the task/context/requirements/output/verification structure.
- Ask AI to generate code from this prompt and observe whether it actually applied the constraints you wrote.
What you produce
- A complete data context write-up
- A structured analysis prompt
- A record comparing AI's generated code against your context
Practice B: Hallucinated Column Hunt
Ask AI to propose an analysis for the practice data and check whether it references a column or business definition that does not actually exist in the data dictionary.
Files used: Analysis package (practice, with challenges), Data dictionary (Traditional Chinese)
Steps
- Ask AI to describe what analyses this data supports and list the columns it plans to use.
- Cross-check every reference against the data dictionary and find any column AI mentioned that does not actually exist, or whose definition was misread.
- Write down why this "hallucinated column" appeared: was the prompt under-specified, or was a column name easy to misread?
- Propose at least one prompt or process improvement that would prevent hallucinated columns.
What you produce
- A list of hallucinated columns found, with the cross-check evidence
- An analysis of why the hallucination occurred
- At least one reusable prevention method
Practice C: Silent Error Review
The practice data contains an analysis that runs cleanly but produces a wrong result because of mixed versions. Find this silent error and fix it.
Files used: Analysis package (practice, with challenges), Data dictionary (Traditional Chinese)
Steps
- Cross-check record_id against status to see whether not-yet-approved rows (draft, superseded) or a duplicated record_id were counted in an aggregation.
- Reconcile row counts before/after filtering and before/after groupby to find where the row count looks wrong.
- Fix the filter, recompute, and compare the result before and after the fix.
- Write down why this silent error was never caught by Python, and what method you used to find it.
What you produce
- The silent error found, with the reasoning behind it
- A before/after comparison of the result
- The concrete method you used to catch this error
Practice D: Verification Ladder
For an analysis conclusion of your own choosing, leave a verification record at each of the four layers: Data → Code → Method → Claim.
Files used: Analysis package (practice, with challenges), Data dictionary (Traditional Chinese)
Steps
- Data layer: confirm the time coverage and known limitations of the source data.
- Code layer: confirm the filter, grouping, and aggregation logic matches the analysis need.
- Method layer: judge whether the chosen statistic (mean or median) fits this group's distribution.
- Claim layer: check whether the final written conclusion faithfully reflects the code output without overstating the evidence.
What you produce
- One concrete verification record per layer
- At least one gap found and fixed during verification
- A comparison of the final claim against the code output
Practice E: Prompt / Code Provenance Log
Build a complete provenance log for the parts of Practice A through D where you used AI assistance.
Files used: Analysis package (practice, with challenges), Independent Lab starter notebook
Steps
- Record the environment (Python and key package versions) and the data version used (file name and source_version).
- Keep the full text of every prompt you sent to AI, along with AI's original code response.
- Record what you changed in the AI-generated code, and why.
- Compile the final code version used and the numbers it produced, and confirm it can be re-run.
What you produce
- A complete provenance log
- A description of the diff between AI's raw output and your edited version
- Confirmation that someone else can re-run it and get the same result
Practice F: Sensitive Data Decision
For the rows in the practice data where sensitive_flag is true, decide whether that data may be shared with an external AI tool, and explain how you handled it.
Files used: Analysis package (practice, with challenges), Data dictionary (Traditional Chinese)
Steps
- Find every row where sensitive_flag is true and describe what kind of information needing special handling it might involve.
- Judge whether this analysis actually needs the detail rows, or only aggregated statistics.
- Choose a handling approach (exclude, aggregate, or de-identify) and explain your reasoning.
- Write down the risk of sharing these rows with an external AI tool without authorization.
What you produce
- A list of rows where sensitive_flag is true, with the reasoning
- Your chosen handling approach and rationale
- A statement of the risk of sharing sensitive data without authorization
The done mark is stored in this browser only (and synced once you join a cohort). Keep your work in your own files or notebook.