Original, traceable sources
Model information comes directly from the public Hugging Face Models API, including model IDs, tags, model-card metadata, base models, licenses, downloads and modification dates. Each profile links to the source repository and README model card.
Model discovery
The collector searches for abliterated, abliteration, uncensored, heretic and obliterated, plus the abliterated tag. Every query follows API pagination to the end, then deduplicates by full repository ID. Keyword discovery can miss untagged models or include results unrelated to text generation.
Model grouping
Explicit base-model relationships within the collected corpus take priority when grouping quantized variants. Other repositories are grouped only within the same publisher after removing recognized packaging suffixes. Names alone never merge different publishers. This conservative approach can retain duplicates and produce different totals from other indexes.
Inclusion threshold and statistics
Model groups need at least 5,000 combined repository downloads in the last 30 days to receive a profile. We scanned 24,238 unique repositories and published 806 models. Downloads and likes are summed within groups, not unique-user counts or lifetime downloads.
Updates and failure protection
GitHub Actions checks every six hours, saves a versioned snapshot after a successful collection, and triggers a Vercel rebuild. Scheduling and builds can be delayed; updates are not instantaneous. Pagination failures, empty results or a drop of more than 35% abort the run and preserve the previous snapshot. Pages show the snapshot time and flag data older than 36 hours.
Unverified metadata
Methods are inferred from names and repository tags. Parameter counts use safetensors metadata when available, otherwise the model name, with that inference disclosed. Missing fields stay unknown. VRAM figures are estimates, not measurements. We do not report refusal rates, capability benchmarks or commercial licensing conclusions.
Reference and independent implementation
The information architecture draws on the public description of Abliterated Models. This site uses independently written code and content, and collects metadata directly from upstream. It does not copy the reference dataset, branding or articles.