Show HN: AI‑curated actual profanity list (github.com)

Very Bad Words (VBW) is an open-source, AI‑curated multilingual profanity wordlist designed to capture actual strong abusive language while avoiding the over‑inclusiveness of many GitHub lists (e.g., blocking common names or mild terms). The creator aggregates multiple sources, runs the combined set through a DistilBERT multilingual profanity classifier (mangalathkedar/profanity-detector‑distilbert‑multilingual) to score candidates, then uses a frontier LLM (Gemini 2.5 Flash) for a second pass and finally human review. The result (vbw.csv) aims to be a lightweight, conservative blacklist for short strings like usernames where a full classifier is overkill but naïve wordlists cause false positives. Technically, VBW’s pipeline is: aggregate sources → classifier score filter → LLM review → human curation. The repo is MIT‑licensed and runnable locally: add your Gemini API key to api_key.txt, pip install transformers torch google-generativeai, then run py aggregate.py, py classify.py, py review.py to reproduce the set. The author emphasizes limits: wordlist filtering is inherently brittle (context, substrings, and location matter), so for longer texts or messages a full classifier model is still recommended. VBW is useful for teams needing a conservative, verifiable blacklist that minimizes collateral blocking while surfacing genuinely abusive terms.