Supplementary data for "Can AI Help German Teachers Select Level-Appropriate Texts? A Cautious Framework for CEFR Estimation"

This dataset supports a small exploratory pilot that compared the CEFR-level estimates of three publicly accessible AI models (Claude Opus 4.8, GPT-5.5 Pro, and Llama 4 Scout) for 14 reading texts drawn from 7 German-as-a-foreign-language (DaF) coursebooks at levels A1, B1, and C1, published between 2021 and 2024 in Germany and Türkiye. Each text was queried once per model under identical zero-shot conditions using a single prompt (provided in English and German).

Publisher CEFR labels are used as curricular reference points, not as independently validated gold standards.The package contains a codebook, a per-text coding table (anonymized metadata and each model's CEFR estimate, with agreement automatically computed against the publisher label), a summary of overall and by-level agreement, and the exact prompt and preprocessing conditions. Full coursebook passages are not included, as they are under copyright; the specific texts can be identified from the metadata and made available to editors and reviewers on request.