I scraped four 4chan boards using the public API, filtered posts through a lexicon of mystical and metaphysical terms, and cleaned the resulting text. Posts were categorized into conflictual, revelatory, and resignationist eschatologies using keyword-based heuristics. I generated embeddings and UMAP projections to visualize thematic clusters, then fine-tuned three GPT-Neo 125M models: one for each eschatology. These models were deployed via a FastAPI backend on HuggingFace Spaces and are accessed through this site’s interface.
More information about my project, as well as access to full datasets and models, can be found on github.