Raw Results for BLADE: A Leakage-Safe Benchmarking Framework for LLM-Driven AutoML

Zenodo (CERN European Organization for Nuclear Research)

Abstract

This record contains raw experiment outputs for the paper "BLADE: A Leakage-Safe Benchmarking Framework for LLM-Driven AutoML". The archive includes experiment logs, generated candidate programs, prompt traces, fitness traces, aggregation outputs, and raw artifacts used to produce the reported tables and figures. The experiments cover three OpenML classification tasks: 146818 (Australian), 359983 (Adult), and 359990 (MiniBooNE). The archive also includes results for the prompt-information ablation comparing none, approx, and exact task-description modes. The data are provided to support reproducibility and inspection of the LLM-driven AutoML case study.

Authors 1

1. Sofoklis Kitharidis corresponding Aachen

   Leiden University

   Affiliation as printed

   Leiden University