Dataset and code for "Generating unplugged lessons with local LLMs: automatic verification and the reliability of model explanations. A feasibility study"
Year: 2026
Type: dataset
Abstract: Code, prompts, rubric, generated lessons and scoring reports for an exploratory feasibility study within the project “Sovereign local AI for educational materials”. Five open-weight models run locally with Ollama (gemma4:e4b and qwen3:8b on a laptop; gemma4:31b, qwen3:32b and gpt-oss:120b on an A100 80 GB GPU) generated a 60-minute unplugged lesson on lossless text compression, adapted from a Code.org/CodeAI Hour of Code activity. A “prompt” mode and a “hybrid” mode with script-based verification are compared, and the reliability of the explanations the models give of their own worked example is analysed. Not peer reviewed; scoring used an LLM-applied rubric, with partial human validation described in the report (its data are not included). The accompanying report is available at https://doi.org/10.5281/zenodo.24345209. The original lesson material is © CodeAI (formerly Code.org) and is not redistributed; see ATRIBUCION.md. Version 0.2.0 adds an analysis of the reliability of the explanations the models give of their own worked example (analisis_xai/). Code, prompts and results are unchanged from version 0.1.0.
Author: Avello-Martínez, Raidell
Institution: Universidade da Coruña
Language: English
Subfield: Education
Field: Social Sciences
Domain: Social Sciences
SDG: Quality education
Open Access status: green
OpenAlex ID: W7234387798
