Rubric dependent associations between Qwen3 model size and Korean constructed response scoring — Code and aggregate reproducibility materials

Abstract

Version 1.0.0 of the computational materials accompanying the manuscript Rubric dependent associations between Qwen3 model size and Korean constructed response scoring by Taeryeong Kim. The package contains analysis and scoring source, frozen configurations and a prompt template, synthetic tests, non-identifying model-by-criterion summaries, paired aggregate bootstrap draws, and scripts reproducing six tables, two figures, and noninferiority sensitivity analyses. It supports offline aggregate replay and computational verification, not an independent reconstruction of historical API calls. Full record-level reanalysis requires separately authorized AI Hub datasets 71818 and 71819 and matching private predictions. Student responses, response identifiers, row-level human ratings and predictions, raw API logs, source rubric text, credentials, expert forms, and private ethics documents are excluded. Evaluation used 4,000 constructed responses and 2,000 essays from the available Validation splits. JDI interpretation and outcome-known diagnostics are explicitly exploratory. This is a computational software release, not publication of the article or a new raw student dataset. Licensing is file-scoped: original author-owned Python code is MIT; author-owned documentation and aggregate results are CC BY 4.0. Dependencies and other third-party materials retain their original terms. See the included license notices and README. No external research funding was received. Substantial AI assistance in design, code, checking and writing is disclosed; responsibility remains with the human author.