nspam - a lightweight model to fight spam
I trained a model from scratch to classify spam on nostr, early indication shows it catches 97% of spam with just a 1mb weight file and can do inference sub 100ms on a pixel 7 with kotlin. (Should work in any language)
Will be shipping in #wisp this week, feel free to try it yourself if you're building a client. I recommend only using it when rendering replies from non followed pubkeys
Model
https://huggingface.co/barrydeen/nspam
Dataset
https://huggingface.co/datasets/barrydeen/nspam-corpus
utxo the webmaster 🧑💻
npub1ut…r50e8
2026-04-15 23:27:55 UTC
Author Public Key
npub1utx00neqgqln72j22kej3ux7803c2k986henvvha4thuwfkper4s7r50e8Published at
2026-04-15 23:27:55 UTCEvent JSON
{
"id": "00008e7823d12c962650436607f60f95b832d57a3313dc4c05b70bf528894189",
"pubkey": "e2ccf7cf20403f3f2a4a55b328f0de3be38558a7d5f33632fdaaefc726c1c8eb",
"created_at": 1776295675,
"kind": 1,
"tags": [
[
"t",
"wisp"
],
[
"client",
"Wisp"
],
[
"nonce",
"2261",
"16"
]
],
"content": "nspam - a lightweight model to fight spam\n\nI trained a model from scratch to classify spam on nostr, early indication shows it catches 97% of spam with just a 1mb weight file and can do inference sub 100ms on a pixel 7 with kotlin. (Should work in any language)\n\nWill be shipping in #wisp this week, feel free to try it yourself if you're building a client. I recommend only using it when rendering replies from non followed pubkeys\n\nModel\nhttps://huggingface.co/barrydeen/nspam\n\nDataset\nhttps://huggingface.co/datasets/barrydeen/nspam-corpus",
"sig": "7d43ec33ca3a01debe4924960dd83e5b7387b18d963c8fc26b25041e8d2f01287ccf791fc462eabc72eacbbb4e05ad2c874f2dffbfff555df6f056fefaba6180"
}