Applesauce

Appearance

Spam Filtering ​

applesauce-extra includes a port of nspam, an on-device classifier for Nostr reply spam. It is a small gradient-boosted tree model, so it runs in plain JavaScript without an ML runtime and scores an author in about a millisecond.

nspam scores authors, not single notes. It reads up to 10 of an author's newest kind 1 replies (events with e tags) and returns the probability that the author is a reply spammer.

Loading the classifier ​

NspamClassifier.load() downloads the model (about 9 MB) from a pinned Hugging Face commit and parses it. Load it once and share it.

ts
import { NspamClassifier } from "applesauce-extra/nspam";

const nspam$ = defer(() => NspamClassifier.load()).pipe(shareReplay(1));

To host or cache the files yourself, pass a url to the directory containing model.txt and calibration.npz, or call parseNspamModel(modelText, calibration) with files you loaded.

Filtering a timeline ​

Pipe a timeline through filterSpam. Every reply in the timeline is added to its author's bundle before filtering, so authors are judged on all the replies the classifier has seen.

ts
const replies$ = eventStore.timeline({ kinds: [1], "#e": [root.id] }).pipe(filterSpam(nspam));

Events that are not kind 1 replies pass through untouched. For single events, use the predicate:

ts
if (nspam.isSpam(event)) return null;

Example: Hide Spam Replies

Scores and thresholds ​

getScore(pubkey) returns { raw, score } for an author, where score is the calibrated probability. Authors at or above threshold (default 0.5) are spam.

ts
const nspam = new NspamClassifier(model, { threshold: 0.8, maxAuthors: 5000 });

nspam.getScore(pubkey)?.score; // 0.97
nspam.score(replies); // score a bundle without remembering it

The classifier keeps bundles for the most recent maxAuthors authors (default 10000).

Integration ​

Combine the score with the viewer's own signals instead of hard-blocking on it. Skip classification for people the user follows and still apply mutes:

ts
const replies$ = combineLatest([eventStore.timeline(filter), user.contacts$]).pipe(
  map(([events, contacts]) => events.filter((e) => contacts?.some((c) => c.pubkey === e.pubkey) || !nspam.isSpam(e))),
);

Best Practices ​

  • Load the model once at module level, it is 9 MB and the classifier is stateful
  • Collapse or dim flagged replies rather than deleting them, the model has false positives
  • Expect weaker scores for authors with fewer than 3 replies and for non-English text
  • Only score kind 1 replies, nspam is not trained on root notes or other kinds