Verifying LLM-Generated Data-Pipeline Transformations: A Hybrid Static–Semantic Approach for Leakage and Data-Quality Faults

Veena G

doi:10.36948/ijfmr.2026.v08i03.81173

Verifying LLM-Generated Data-Pipeline Transformations: A Hybrid Static–Semantic Approach for Leakage and Data-Quality Faults

Author(s)	Ms. Veena G
Country	India
Abstract	Large language models (LLMs) are increasingly used to generate data-pipeline and feature engineering transformations from natural-language specifications. This introduces a new and under-examined risk surface: transformations that are syntactically valid and schema-correct yet semantically wrong, silently injecting temporal leakage, granularity errors, join fan-out, silent row loss, and data-quality defects. Existing text-to-SQL verifiers target the correctness of a single natural-language-to-SQL query and do not address transformation- and pipeline-level data-flow faults, which today are caught only reactively—after a model trained on a leaking feature posts implausible offline accuracy. We make three contributions. First, we define a six-category fault taxonomy for generated transformations grounded in real feature-engineering failure modes. Second, we propose a lightweight hybrid verifier that combines static dataflow analysis with semantic execution checks, including a gold-free metamorphic leakage test: removing future rows must not change any historical feature value. Third, we build a fault-injection benchmark on synthetic and real (Rossmann Store Sales) data and evaluate against a static validation baseline and a strong LLM-as-judge baseline. The hybrid verifier attains perfect precision (no false alarms), is deterministic and inexpensive, and detects structural, temporal, and granularity faults; a strong LLM-as-judge attains higher overall recall on semantically subtle faults but over-flags correct transformations and is blind to granularity errors. The two approaches are complementary, motivating combined verification for trustworthy LLM-generated data pipelines.
Keywords	data engineering, feature engineering, data leakage, LLM code generation, software verification, data quality, trustworthy AI.
Field	Computer > Artificial Intelligence / Simulation / Virtual Reality
Published In	Volume 8, Issue 3, May-June 2026
Published On	2026-06-13
DOI	https://doi.org/10.36948/ijfmr.2026.v08i03.81173

View / Download PDF File

About IJFMR Fees & Payment Current Issue Publication Archive	Submit Research Paper Track Submission Status Publication Guidelines Publication Ethics Peer Review & Plagiarism	Join as a Reviewer Editors & Reviewers Reviewer Referral Program Get Reviewer Membership Certi.	Website/Journal Policies Usage Policy Content Policies Privacy Policy

Contact Us		+91-9687-828-838	editor@ijfmr.com

International Journal For Multidisciplinary Research

A Widely Indexed Open Access Peer Reviewed Multidisciplinary Bi-monthly Scholarly International Journal

Verifying LLM-Generated Data-Pipeline Transformations: A Hybrid Static–Semantic Approach for Leakage and Data-Quality Faults

Share this