Replace toxicity model with multi-label classifier

This commit is contained in:
Jordan Wages 2026-08-26 17:00:49 -05:00
commit 1568c787f8
24 changed files with 2581129 additions and 119717 deletions

View file

@ -1,14 +1,16 @@
import { afterEach, describe, expect, it } from "vitest";
import { applyResult, removeOrphanedPlaceholders, restore } from "../src/content/filter";
import { applyResult, removeOrphanedPlaceholders, restore, selectFilterLabel } from "../src/content/filter";
import { DEFAULT_SETTINGS } from "../src/shared/types";
const scores = { toxicity: .91, severe_toxicity: .1, obscene: .3, threat: .2, insult: .8, identity_attack: .1, sexual_explicit: .1 };
describe("content filtering", () => {
afterEach(() => document.body.replaceChildren());
it("collapses toxic content and restores it", () => {
const element = document.createElement("article");
document.body.append(element);
applyResult(element, { requestId: "1", id: "1", textHash: "x", label: "toxic", probability: .91, navigationId: "n" }, DEFAULT_SETTINGS);
applyResult(element, { requestId: "1", id: "1", textHash: "x", scores, navigationId: "n" }, DEFAULT_SETTINGS);
expect(element.style.display).toBe("none");
expect(document.querySelector("[data-vibeguard-placeholder]")).not.toBeNull();
restore(element);
@ -19,13 +21,19 @@ describe("content filtering", () => {
it("preserves the original display value when filtering is reconciled repeatedly", () => {
const element = document.createElement("article");
document.body.append(element);
const result = { requestId: "1", id: "1", textHash: "x", label: "toxic" as const, probability: .91, navigationId: "n" };
const result = { requestId: "1", id: "1", textHash: "x", scores, navigationId: "n" };
applyResult(element, result, DEFAULT_SETTINGS);
applyResult(element, result, DEFAULT_SETTINGS);
restore(element);
expect(element.style.display).toBe("");
});
it("uses the highest enabled category above the threshold", () => {
const result = { requestId: "1", id: "1", textHash: "x", scores, navigationId: "n" };
expect(selectFilterLabel(result, { ...DEFAULT_SETTINGS, enabledLabels: ["insult", "threat"], threshold: .5 })).toEqual({ label: "insult", probability: .8 });
expect(selectFilterLabel(result, { ...DEFAULT_SETTINGS, enabledLabels: [], threshold: .5 })).toBeUndefined();
});
it("removes placeholders no longer paired with the current post element", () => {
const stale = document.createElement("div");
const orphan = document.createElement("div");