Sociocultural Mediation Through Large Language Models: Enhancing Learner Agency and Identity in Chinese EFL Speaking Contexts

Anonymized dataset and analysis materials supporting "Sociocultural Mediation Through Large Language Models: Enhancing Learner Agency and Identity in Chinese EFL Speaking Contexts." This dataset covers the same 8-week LLM-based speaking intervention (October-December 2025) as the companion dataset for "Large Language Models, Speaking Anxiety, and Self-Efficacy in Chinese EFL" (see linked DOI below), analyzed here through a Sociocultural Theory lens with a QUAL-dominant embedded mixed-methods design. Contents: anonymized full interview transcripts and individual transcript files for 12 semi-structured interviews (437 speaker turns); the complete qualitative coding trail (initial codes, human coding, and the 434-excerpt master coding sheet with theme assignments); the APA theme summary table; raw and processed online student engagement scale (OSE) responses; main- and second-rater voice recording task (VRT) proficiency scores (pre- and post); the weekly AI practice feedback log (196 responses, tool usage, and comfort ratings); the correlation matrix and Spearman correlation outputs; and the anonymized participant consent records. Data quality notes: A duplicate participant submission (C-16) in the OSE response file has been identified and corrected in the processed files; the corrected N is 53 unique participants (not 55, an earlier miscount) with 27 unique OSE respondents (not 28). Two source-ID entries flagged during data cleaning as corrupted in the original export (raw values "20" and "356") have been resolved to C-20 and treated as unmapped, respectively; see the coding sheet's Data_Quality_Flag column for full detail. The qualitative coding was conducted by a single researcher, cross-checked against participant feedback data rather than a second independent coder; this is disclosed as a limitation in the associated manuscript. Shared-data disclosure: This dataset shares its source intervention with the dataset accompanying "Large Language Models, Speaking Anxiety, and Self-Efficacy in Chinese EFL" (currently under review at Language Learning & Technology). The two studies apply distinct theoretical frameworks (sociocultural theory here; foreign language anxiety theory and Bandura's self-efficacy framework there) to non-overlapping research questions and different subsets of the underlying data, as disclosed in both manuscripts' cover letters and methods sections. Citation: Please cite the associated paper and this dataset DOI if you use this data. Re-users must not attempt to re-identify participants.