Core Concepts

Dataset Review Engine

Automated code review discipline for tabular datasets

In traditional software engineering, developers submit pull requests and run automated linters and code reviews before merging code to production. In data engineering and machine learning, datasets are frequently trained on without any formal review step — leading to silent model failures in production.

Featuresmith's Dataset Review Engine establishes code review discipline for tabular datasets by running 10 specialized reviewers to evaluate schema health, data types, missingness, duplicates, constant columns, cardinality, distributions, feature quality, and target leakage.

Reviewers vs. Rules

Rules (Atomic Assertions): Atomic checks evaluated by the Rule Engine (e.g. "Are missing values in column X greater than 20%?"). Rules produce raw RuleFinding objects.
Reviewers (Section Aggregators): Higher-level domain inspectors. Each reviewer evaluates one aspect of dataset health (e.g. MissingValueReviewer), aggregates related rule findings, assigns a section severity, and compiles a clean ReviewSection.

Review Output Structure

Invoking fs.review(dataset) returns a single frozen ReviewResult dataclass containing:

sections: List of ReviewSection objects (one per active reviewer).
overall_summary: Human-readable text summary of overall evaluation results.
score: The 0–100 MLReadinessScore (or None if scoring is disabled).