Replication Package for 'Understanding Open-Source AI Adoption: What Practitioners Discuss, Where Support Breaks Down, and Why Deployment Fails'

Replication Package: The Friction of Open-Source AI Adoption Submission ID: 6916 Overview This package contains the complete dataset and analytical pipeline used to examine 24,000 discussion threads from the Hugging Face ecosystem (2020–2025). The study identifies a "Semantic Drift" from theoretical inquiry to infrastructure engineering and documents a "Compute Divide" where hardware constraints act as a primary barrier to model adoption. Repository Structure code/: Contains the Jupyter notebooks required to replicate each phase of the study. HFForumExtraction.ipynb: The asynchronous scraper used for initial data acquisition. Validation.ipynb: Scripts for calculating Inter-Rater Reliability (IRR) between human annotators and the LLM agent. Analysis_RQ1.ipynb: Quantitative tracking of semantic shifts over time. Analysis_RQ2.ipynb: Analysis of the "Support Gap" and topic resolution rates. Analysis_RQ3.ipynb: LLM-augmented root cause analysis and statistical modeling of adoption friction. hf_forum_data/: The primary datasets. hf_forum_dataset.parquet: The "Gold Master" corpus of ~24k threads. hf_forum_autopsy_ultimate.parquet: The subset of 3,500 unsolved threads annotated for root cause analysis. Quick Start Environment: Ensure Python 3.9+ is installed. Dependencies: Install required libraries: pandas, numpy, matplotlib, seaborn, scikit-learn, scipy, statsmodels, aiohttp, google-generativeai, and pyarrow. Execution: The notebooks are designed to be run sequentially. While the raw data is provided, Analysis_RQ3.ipynb includes the LLM annotation logic which requires a Google Gemini API key to run from scratch. However, the provided autopsy dataset allows for full replication of the statistical results without further API costs. 📜 License All code is provided under the MIT License. The dataset is intended for academic research purposes.