FELA · byte-level
Fast Toxicity Ranking
Setup
Live ranking a stream of comments on two signals in one pass, toxicity and personal info (PII) detection.
Pros
Runs offline in the client · Prevent custodying toxic content!
Usage
Watch the live triage queue rerank itself, pause or resume the stream, or drop in your own comment.
This thing does WHAT NOW?
Here we got a compact 12.7 million parameter FELA model reading raw bytes (look Ma! NO TOKENIZERS!)
and scoring each comment for toxicity, then keeping a queue sorted so the worst rises to the top. Yes... it is
tokenizer free, meaning it could support text in any language - treating it as just a byte sequence, with numerous other nerdy advantages. The point of the demo is
the ordering and the speed: it triages a stream of comments on a single CPU
core, no GPU and no server round trip. You can imagine preventing hate speech from spreading at the speed of love!
Why does FELA keep winning?
The encoder here is a bidirectional FNO (global Fourier mixing, O(N log N))
interleaved with gated linear attention (O(N), constant memory), with no
quadratic attention matrix. In plainer terms - we looked at the stinky, expensive parts of
modern AI and decided "yeah we don't need that stuff." That is what lets a full toxicity pass run in milliseconds
on one CPU core. Triaging a stream of comments costs cents instead of a per call API
bill - and high confidence detections could prevent abuse.
Usage
See our Github repo here containing a very convenient, drop in,
self moderating modern JS/TS/React text input widget. Give us a Github star if you can! And consider sponsoring our FOSS work!
Full card on Hugging Face ↗