English Learning Behaviour Log Dataset for Knowledge Tracing and Adaptive Guidance

Each answer carries on average 0.3495 hint requests, 0.2071 option switches and 0.2902 answer changes. Learners are stratified by sequence length into a short layer of fewer than 20 answers covering 18.58 percent of learners, a medium layer of 20 to 80 answers covering 57.33 percent and a long layer of more than 80 answers covering 24.08 percent. The learners are split into 840 training, 120 validation and 240 test learners, and each learner belongs to one split only.<br><br>The release contains four files. english_learning_events.csv is the event table in UTF-8 with a header row. dataset_summary.json reports the counts, the per dimension statistics, the split sizes, the sequence length layers and the Bayesian ceiling of predictive performance, which is 0.9013 on the short layer, 0.9058 on the medium layer, 0.9101 on the long layer and 0.9079 overall.

DATASET.md documents the field dictionary, the generative process, the statistics and the limitations. LICENSE.md states the data licence.<br><br>The records were produced by a documented stochastic generative process, implemented in src/data/generate_dataset.py of the companion software release, and the released seed reproduces the file bit for bit. Field granularity, event types and the value ranges of the reported statistics follow the record structure of public English learning and practice log datasets, among them EdNet, the ASSISTments skill builder logs and the Duolingo second language acquisition corpus, which are listed in the References field.

A reproducible simulation is used because the four behaviour attributes the method consumes, hint requests, option switches, answer changes and answering time, are not jointly recorded in any single public log, and a common field set across every compared model cannot otherwise be guaranteed. The generative process is stated in full in DATASET.md and in the released generator, so every statistic in the companion manuscript can be regenerated from source.<br><br>Released for research use under CC BY 4.0.</pre><p></p>