It's fascinating that there are lightweight models that can achieve 'real time' or at least 'near real time' under 100ms.

One of those models is from detoxify (unbiased-small) which achieves evaluation score of 93.74 (Top Performer score: 94.73).

nostr:nevent1qqs9xr3aegytyds0fmlwz0x2pzyaafccedxy9emjrj5589ltwwwcuyspp4mhxue69uhkummn9ekx7mqzyq4k0eyqklue62p4dp9g7unkmphdh68rrr49tnmhen74t8zlynmy2qcyqqqqqqg6udx7n

Update:

There were 4358 sample data classified within 24 hours with minimum classification thresold score of 0.2.

Note:

Early WIP Version

Update:

There were 4358 sample data classified as potential toxic notes/posts from 75000+ total notes/posts within 24 hours, with minimum classification thresold score of 0.2.

I will take a look manually to see how the data behave. Hopefully, it won't 'poison' much to read them 😅

nostr:nevent1qqs9xr3aegytyds0fmlwz0x2pzyaafccedxy9emjrj5589ltwwwcuyspp4mhxue69uhkummn9ekx7mqzyq4k0eyqklue62p4dp9g7unkmphdh68rrr49tnmhen74t8zlynmy2qcyqqqqqqg6udx7n