Dataset-Level Reliability Assessment of LLM-Extracted Metallurgical Data: High-Chromium Steels Corpus

This dataset contains structured experimental information extracted from 23,861 full-text journal articles (2005–2025) on high-chromium steels (8–20 wt.% Cr) using a staged large language model (LLM) workflow. The corpus was retrieved in XML format via the Elsevier API and filtered for domain relevance prior to preprocessing and extraction. The dataset comprises three primary data domains: Chemical composition (CC): alloy compositions reported in wt.%. Heat treatment (HT): processing parameters including temperatures, times, and treatment sequences. Phase-related data (PS): secondary phase identity and reported particle sizes. Extraction was performed through a multi-stage pipeline including: (i) LLM-based text classification, (ii) structured data extraction, and (iii) automated LLM-based review for error characterisation. The review stage labels extracted entries according to three error categories: Fabricated Data (FD) Missing Data (MD) Swapped Data (SD) Entries flagged during automated review are retained with diagnostic labels, enabling downstream risk-aware filtering rather than binary acceptance/rejection. Post-review filtering yielded 36,658 composition-related data points from 19,539 unique articles. Within the 8–20 wt.% Cr subset, 12,399 heat-treatment records and 11,895 phase-related entries are included. Targeted manual sampling indicates residual agreement rates of 97.86% (CC), 96.78% (HT), and 85.31% (PS), with remaining errors dominated by phase-scale misclassification and fabrication in semantically ambiguous context. This dataset is not presented as error-free. Instead, it provides structured metallurgical data together with explicit uncertainty characterisation, enabling quantitative assessment of residual risk prior to downstream statistical modelling, alloy design studies, or meta-analysis. The repository also links to the extraction and evaluation code used to generate and review the dataset, supporting methodological transparency and reproducibility.