Wetland Conservation in Nebraska Local Plans: Planning Documents, Human Reference Scores, and LLM Evaluation Records

Description

This dataset accompanies the study “Using ChatGPT to Evaluate the Integration Status of Wetland Conservation in Local Plans” and contains archived materials for evaluating how wetland conservation is addressed in local comprehensive plans in Nebraska, USA. The planning documents were obtained from publicly accessible government websites. The study corpus comprises 25 plan units, with five used for prompt development and 20 for validation. Human reference scores cover 56 indicators per plan, giving 1,400 plan–indicator scores on an ordinal scale: 0 = not addressed, 1 = partially addressed, and 2 = fully addressed. The validation prediction table contains 4,480 scheduled evaluations across GPT-4o and GPT-4o mini under basic and optimized prompts. Fifteen model scores are unavailable and remain explicitly missing. Binary scores are post-hoc recodings of the ternary scores, with 0 indicating absence and 1–2 indicating presence. The package includes archived planning PDFs and a checksum manifest; a plan inventory and development/validation assignments; human reference labels; model predictions, requests, responses, and explanations; prompt and scoring-framework materials; optimization-history records; and statistical CSV tables with an offline analysis script. The resolved model-output archives cover 40 historical plan units, whereas the study validation analyses use only the 20 designated validation plans. The README documents file contents, variables, missing values, and provenance limitations, including three scores retained in the prediction table but absent from the supplied resolved JSON files. The model scores, explanations, and optimization text are AI/LLM-generated experimental outputs. The study code is shared on GitHub at https://github.com/HaishanLinHZNU/LLM_Plan_Evaluation. A local historical code-and-results snapshot is also included in this package. The offline analysis uses the supplied archived prediction table and requires no model API calls.

Institutions

• Hangzhou Normal University

  Zhejiang, Hangzhou

Categories

Licence