idta-smt-extraction-prompts

Zenodo (CERN European Organization for Nuclear Research)

Abstract

Five SHA-pinned LLM prompts (v0–v4) for schema-aware constraint extraction from IDTA Submodel Template PDFs. Each version adds exactly one design feature on top of the previous (layout-aware reading, bounded template node graph, honesty contract, few-shot examples), making the v0→v4 series a controlled single-axis ablation. The prompts emit JSON arrays of typed constraint records with calibrated abstention via a closed status vocabulary. Headline F1 on IDTA 02006 Digital Nameplate v3.0.1 with qwen2.5:7b-instruct at temperature 0.0: v2 F1=0.481 at precision 0.905, v3 precision 1.000 at F1=0.474. Within-prompt- SHA stability rate 1.000 across K=5 runs.

Authors 1

1. Mohammad Hadi Khori Jannatabadi corresponding Aachen Chair of Information and Automation Systems

   RWTH Aachen University

   Affiliation as printed

   Chair of Information and Automation Systems, RWTH Aachen University