Dataset and Analysis Scripts for: An 8-Week LLM Speaking Intervention and EFL Anxiety: A Mixed-Methods Pilot

Abstract

This dataset accompanies a mixed-methods pilot study of an eight-week intervention (22 October – 17 December 2025). In the intervention, 54 university EFL students in China used freely available large language model chatbots for low-stakes speaking practice. The chatbots used were DeepSeek, Doubao, ChatGPT, Kimi Chat and others, with weekly prompts shared via WeChat. The dataset contains: Quantitative data: raw Microsoft Forms exports and processed scores for: the Foreign Language Classroom Anxiety Scale (15 items), before and after the intervention the English Speaking Self-Efficacy Scale (10 items), before and after the intervention speaking rubric scores (VRT) a researcher-developed engagement questionnaire a weekly AI-practice log with 196 entries Qualitative data: anonymised transcripts of semi-structured interviews with 12 participants, preliminary computational coding files, and the final thematic framework. Code and results: a single Python script (Code_and_Scripts/analysis_pipeline.py). It scores every instrument, applies documented exclusion rules, and reproduces all reported statistics, tables and figures. It also produces a sensitivity analysis. Start with Documentation/README.txt, which explains scoring, exclusions, file contents and anonymisation. Participants are identified only by codes. Audio recordings are not shared. What's new in Version 3 (supersedes all earlier versions): Scoring bug fixed. Some Likert responses were stored with a non-breaking space or different capitalisation, and earlier versions summed these as zero. One FLCAS item was also dropped. All FLCAS statistics from earlier versions are superseded. OSE scoring corrected. Negatively worded engagement items are now reverse-scored. Exclusions documented. Eleven post-test responses time-stamped before the intervention began are now excluded, as are unlinkable IDs, incomplete responses and one duplicate. A sensitivity analysis that includes them is provided. Tool-usage figures corrected. Anonymisation improved. Place and institution names in the transcripts were replaced with placeholders, and one personal name was removed from a raw file. Documentation rewritten. Legacy scripts are kept but marked as deprecated.