Anonymized Dataset and Scripts for LLM Intervention in Chinese EFL: Reducing FLA and Enhancing Proficiency

This dataset supports the study titled "The Role of Large Language Models in Reducing Foreign Language Anxiety and Enhancing English Proficiency for Chinese EFL Learners." The research employed a sequential explanatory mixed-methods design to evaluate an 8-week asynchronous speaking intervention using large language models (LLMs) such as DeepSeek and Kimi Chat) delivered via WeChat assignments to 52 intermediate-level Chinese EFL students at one of the Chinese university.The intervention aimed to provide low-stakes, non-judgmental speaking practice to address foreign language anxiety (FLA), increase speaking self-efficacy, and improve oral proficiency in a high-anxiety sociocultural context characterized by Confucian heritage values (e.g., fear of losing face) and exam-oriented education.Quantitative findings showed significant reductions in FLA (FLCAS: t(44) = -3.964, p < .001, d = 0.59), increases in self-efficacy (adapted ESSES: Wilcoxon signed-rank test, p < .001, r = 0.63), and gains in proficiency (VRT rubric: t(11) = 4.011, p = .002, d = 1.16), with engagement (OSE) positively correlated with proficiency gains (r = .682, p < .05). Qualitative thematic analysis of 470 interview turns from 12 participants identified key mechanisms, including non-judgmental AI feedback, cultural relevance of prompts, and risks of over-reliance.Contents of the dataset Quantitative_Data: Anonymized pre- and post-intervention scores for FLCAS, ESSES, OSE, and VRT; processed CSV files; master merged dataset.Qualitative_Data: Anonymized interview transcripts (full and individual speaker files using C-XX codes); initial coding and final refined themes (CSV).Code_and_Scripts: Python scripts for data cleaning, processing, statistical analysis (paired t-tests, Wilcoxon, correlations), thematic coding support, and VRT scoring.Documentation: Intervention summary PDF, anonymized consent form template, README with variable descriptions and replication guide, ethics and anonymization notes. How to replicateAll participant data is fully anonymized: no names, IDs, locations, or identifiable information remain. Transcripts and qualitative files use speaker codes (e.g., C-01 to C-30). Quantitative files use pseudonymized participant codes where applicable.To replicate quantitative results: Load Master_Quantitative_Data_Anonymized.csv and run quant_analysis_final.py or vrt_real_analysis.py for descriptive statistics, paired tests, effect sizes, and correlations.To replicate qualitative analysis: Use Final_Refined_Themes.csv and Initial_Human_Coding_Anonymized.csv together with read_transcripts.py or initial_coding.py for theme verification and LDA support.See README.txt for detailed variable descriptions, folder structure, and step-by-step instructions.This dataset enables full reproduction of the reported statistical results and thematic findings while strictly protecting participant privacy in accordance with informed consent and ethical guidelines.