---

name: dataset-comparison description: "Compare two pandas DataFrames or two subsets of one DataFrame with Sweetviz, including target-aware validation and deterministic HTML output checks." disable-model-invocation: true metadata: disco-role: operating license: MIT

Dataset Comparison

Use this sub-skill when the task asks for Sweetviz comparison, train/test profiling, dataset drift/profile comparison, or subgroup comparison such as "male vs female". It owns sweetviz.compare() and sweetviz.compare_intra() guidance only.

Route here when

• Comparing two existing pandas DataFrames, usually train/test, baseline/current, or reference/candidate.
• Splitting one DataFrame into two named groups with a boolean condition and comparing the groups.
• The comparison includes a target feature and needs validation before rendering.
• The user mentions compare, compare_intra, subgroup split, non-boolean condition, empty split, or compare-only columns.

Route elsewhere

• Single-DataFrame sweetviz.analyze() or render-only/browser/notebook details: use the report-generation sub-skill.
• Detailed FeatureConfig, type forcing, config override files, duplicate-column cleanup, or input preflight beyond comparison-specific checks: use the configuration-and-data-handling sub-skill.
• Install/import/package-data/font issues: use the Sweetviz root troubleshooting reference.

Operating flow

1. Choose the constructor:
   • sweetviz.compare([source_df, "Source name"], [compare_df, "Compare name"], target_feat=..., feat_cfg=..., pairwise_analysis=...) for two DataFrames.
   • sweetviz.compare_intra(source_df, condition_series, ["True group", "False group"], target_feat=..., feat_cfg=..., pairwise_analysis=...) for a split of one DataFrame.
2. Validate comparison inputs before constructing the report:
   • No duplicate columns in either DataFrame.
   • compare_intra condition has the same length as source_df, has plain boolean dtype, and creates non-empty true and false groups.
   • Target feature exists in the source, is not skipped, is numeric or boolean, and has no missing values. If the compare DataFrame also contains the target, it must also have no missing values.
   • Any names used in FeatureConfig are case-sensitive and must be present in the source DataFrame.
3. Prefer explicit dataset names ([df, "Training Data"] or (df, "Training Data")) so report sections are interpretable.
4. Pick pairwise_analysis: use "off" for deterministic or wide smoke runs, "auto" for ordinary small reports, and "on" only when association cost is acceptable.
5. Render with report.show_html(..., open_browser=False) for automated work and validate the output file exists and is non-empty. For layout/browser/notebook choices, defer to report-generation.

Bundled references and helper

• Comparison workflows: train/test and compare_intra recipes with tiny schemas and deterministic output validation.
• API reference: verified signatures, accepted input shapes, naming, target/config interactions, pairwise behavior, and column mismatch notes.
• Troubleshooting: non-boolean conditions, empty splits, duplicate columns, target NaNs, missing FeatureConfig names, and pairwise threshold behavior.
• Smoke script: offline argparse helper that builds tiny fixtures, runs compare and/or compare_intra, writes HTML without opening a browser, and can demonstrate invalid non-boolean conditions.

Non-goals and cautions

• Do not pass verbosity to sweetviz.compare() or sweetviz.compare_intra(); their public signatures do not accept it. If progress output must be controlled, use Sweetviz configuration defaults or the lower-level report class guidance from the configuration/reporting sub-skills.
• Do not use maintainer or network-update helpers for end-user EDA comparison workflows.
• Optional Comet.ml logging is credentialed external behavior owned by reporting guidance; do not require it for comparison verification.