CogTax: dataset of taxonomy-labelled Linux/bash commands

Abstract

585 Linux/bash commands labelled with the four-level CogTax cognitive-operational taxonomy (1-4), used to evaluate whether taxonomy levels are automatically recoverable from command syntax and semantics. Companion dataset for the article "CogTax: A Four-Level Cognitive Taxonomy for Command-Line Computing Education" (PeerJ Computer Science, under review). Composition: 117 human-authored exam model answers (L1: 26, L2: 34, L3: 27, L4: 30) and 468 LLM-generated synthetic commands, exactly balanced across levels (117 per level). The human commands are model answers written by the teaching staff for eight exam models of a second-year Linux/Bash course, each tagged with its taxonomy level by the same staff. Contents: the two command datasets (commands_dataset.parquet, the 585-command evaluation set, and commands_dataset_synthetic.parquet, its 468-command training split), the redacted exam source file, the AST feature extractor (src/ast_extractor.py) used to compute the structural metrics reported in the article, and the scripts used to build the dataset from its sources (redact_exam_models.py, build_questions.py, build_dataset.py). Running `build_dataset.py --check` verifies the published parquet reproduces exactly from these sources. Data protection: student examination responses and grades are not included. The study that produced this corpus also collected 87 student answer sheets, which identify living individuals and are not covered by a consent to publish; no result in the article depends on them. The redacted exam file also has the shared course account password and the internal grading scheme (per-section rubrics and marks) removed, as course-internal material unrelated to this study. Licensing: the dataset files (data/) are released under CC BY 4.0. The accompanying source code (src/, scripts/) is released under the MIT licence; see the LICENSE file included in the archive.