preflight-data 0.1.1

---

Dataset readiness checks for ML pipelines. Use preflight to catch data blockers before training and deployment.

Why preflight

• Runs fast checks for data quality, target risk, schema/type issues, split integrity, and statistical anomalies.
• Produces machine-readable findings and a CI gate decision (PASS/FAIL).
• Keeps output stable with schema versioning for downstream tooling.

Quickstart

Install:

pip install preflight-data

Python API:

import pandas as pd
import preflight

df = pd.read_csv("data.csv")
report = preflight.run(df, target="churn", profile="ci-balanced")

print(report)

CLI:

preflight run data.csv --target churn --profile ci-balanced --format text

Example output:

Preflight Run Report
────────────────────────────────────────────────────────
Gate: PASS
Heuristic Score: 97.0/100
Profile: ci-balanced
Dataset: 1000/1000 rows analyzed across 6 columns
Target: churn
Summary: 8 info, 1 warn, 0 error, 0 critical

Gate reasons:
- No findings met fail conditions for this profile

Findings:
- [WARN] completeness.missingness: Overall missingness is 1.8% (108 cells missing across dataset). (confidence=0.95)
- [INFO] duplicates.exact: No exact duplicate rows detected. (confidence=0.95)
- [INFO] balance.class_imbalance: Class distribution is within configured tolerance. (confidence=0.90)

HTML report preview

If you want to see what the generated HTML report looks like, open:

• preflight_report.html
• preflight_public_datasets_demo.ipynb

The notebook contains rendered report.to_html() output cells.

Core model

• finding: one detected issue or advisory signal
• severity: info | warn | error | critical
• gate: policy decision based on severities (PASS | FAIL)
• score: heuristic summary for trend/comparison, not statistical truth

Score guidance:

• Good for rough trend tracking across runs
• Not a probability of model success

Common workflows

Dataset readiness (single table)

report = preflight.run(df, target="churn", profile="ci-balanced")

Split integrity (train/validation or train/test)

split_report = preflight.run_split(train_df, valid_df, profile="ci-balanced")

Policy profiles

Built-in profiles:

• exploratory: permissive, useful in notebooks
• ci-balanced: practical CI default
• ci-strict: highest sensitivity for blocking conditions

Example:

preflight run data.csv --target churn --profile ci-strict --format json

--fail-on override:

preflight run data.csv --target churn --profile ci-balanced --fail-on error,critical

Policy argument rules:

• Use either --profile or --policy-file (mutually exclusive).
• --fail-on is only supported with --profile.
• Invalid policy/config files fail fast at load time.

CLI reference

# Recommended policy-first commands
preflight run data.csv --target churn --profile ci-balanced --format json
preflight run-split train.csv test.csv --profile ci-balanced --format markdown

# Optional artifacts
preflight run data.csv --target churn --format text --output report.txt --output-html report.html

# Compare against baseline JSON report
preflight compare current.json baseline.json --max-score-drop 3 --fail-on-new-error

# Suppressions
preflight suppress add --file suppressions.json --check-id leakage.high_correlation --reason "known safe"
preflight suppress list --file suppressions.json
preflight suppress validate --file suppressions.json --fail-on-expired

HTML output

CLI:

preflight run data.csv --target churn --profile ci-balanced --format text --output-html report.html

Python:

report = preflight.run(df, target="churn", profile="ci-balanced")
html = report.to_html()
with open("report.html", "w", encoding="utf-8") as f:
    f.write(html)

This creates a shareable HTML report you can attach to CI artifacts, docs, or review tickets.

Exit codes:

• 0: gate pass
• 2: gate fail or explicit CLI validation failure

Output schema contract

RunReport.to_dict() includes stable contract keys:

• schema_version
• run
• dataset
• gate
• score
• summary
• findings

Per-finding payload includes evidence and explainability fields:

• check_id, title, domain, severity, suppressed
• suggested_action, docs_url
• evidence.metrics, evidence.threshold, evidence.samples

Examples

• Realistic workflow notebook:
  • simple_example.ipynb
• Public dataset demo notebook:
  • preflight_public_datasets_demo.ipynb
• Script demo:
  • public_datasets_demo.py

Legacy compatibility

Legacy check(...) and check_split(...) APIs are still available for compatibility, but run(...) and run_split(...) are recommended for policy-first workflows.

Migration status: the policy-first runner now uses native checks for class balance, completeness, leakage, duplicates, distributional health, correlations, and types. Legacy APIs remain supported during migration.

Compatibility namespace:

• preflight.legacy.check(...)
• preflight.legacy.check_split(...)
• preflight.legacy.Report

Development

make env
conda activate preflight
make install-dev
make test
make lint
make typecheck
make build

Supported versions

• Python: 3.9-3.13
• pandas: >=1.3
• numpy: >=1.21

License

Apache-2.0