clvv/hf-uncensored-model-popularity

This repository holds the data and code behind a single chart that someone posted on Twitter. The chart looks at how popular so-called uncensored AI models are on Hugging Face, which is a public site where people share language models. The author is upfront that this is a fully AI-generated analysis, meant as a reproducible public snapshot rather than an official Hugging Face dataset or an audited academic study. The headline number is that uncensored or heretic-tagged model repos created so far in May 2026 account for about 15.24 percent of rolling downloads among language-model-style repos. Across all time the share is smaller, around 0.9 percent of downloads and 1.8 percent of repos, so the May spike is the part being highlighted. The method explains what was counted. The denominator pulls in all Hugging Face model repos tagged with text-generation, text2text-generation, conversational, image-text-to-text, gguf, or llama.cpp. The numerator narrows that group down to repos whose ID or tags include words like uncensored, heretic, abliterated, censorless, jailbreak, or liberated. Download and like counts are point-in-time numbers taken from Hugging Face's public API, grouped by the month each repo was created. The repo ships a Python collector script, a compressed CSV snapshot of the raw data, a smaller monthly aggregate CSV, a self-contained HTML report, the SVG chart files, and a metadata file listing the exact filters and keyword list used. A citation file records hashes and provenance. To reproduce, you install the huggingface_hub Python package and run the collector script. Fresh runs will give slightly different numbers because the Hugging Face data is live, so the included CSVs are the way to verify the exact chart shown in the tweet.